Assignment Introduction Potential Impacts of Artificial Intelligence .pdf

Chapter 1: Introduction to the Study

Introduction

The volume of data in health care is exploding exponentially. This trend is

supported by digitization of health care data from different sources along with advances

in imaging technology and molecular diagnostic methods (Murdoch & Detsky, 2013;

Raghupathi & Raghupathi, 2014). The overwhelming volume of data in health care

contributes to chaos and uncertainty that can lead to diagnostic errors with devastating

consequences such as chronic pain, disability, or death (Gupta et al., 2017; Saber-Tehrani

et al., 2013). The radiologist is rapidly becoming one of the most important curators and

gatekeepers of big data. Despite this important role, the complexity and volume of

available data during the interpretive stage of radiology workflow outpaces the individual

radiologist’s capacity to make fully informed and timely decisions (Thrall et al., 2016).

The rate of diagnostic error associated with the interpretation of abnormal radiology

studies has been reported as high as 30% with retrospective review of diagnostic imaging

studies yielding even higher error rates (Berlin, 2007; Brady et al., 2012; Donald &

Barnard, 2012).

The causes of errors during the interpretive stage of radiology workflow include

exposure to complex data from new imaging technology, limited access to relevant non-

imaging data, increasing workload, physiological fatigue, and human bias (Brady et al.,

2012; Donald & Barnard, 2012; C. S. Lee, Nagy, Weaver, & Newman-Toker, 2012).

Without adequate technological assistance, the human interpretive process within

radiology workflow will become progressively more inaccurate, inefficient, and untimely

(Croskerry 2013; Manrai et al., 2014; Obermeyer & Emanuel, 2016; Ragupathi &

Ragupathi, 2014). The solution requires the adoption and implementation of new data

management and decision support solutions such as artificial intelligence (AI).

AI offers a variety of approaches to assist in problem solving and prediction

(Pannu, 2015). The process encompasses different computational methods such as

machine learning, deep learning, and cognitive computing. Using any of these AI

solutions during radiology workflow might improve diagnostic accuracy and timing. AI

is capable of revealing biological variability, heterogeneity of pathology, and

comorbidities, all of which support more personalized and precise diagnosis (Gillies,

Kinahan, & Hricka, 2016; Limkin et al., 2017; Pannu, 2015; Yip & Aerts, 2017). AI will

fundamentally alter the field of radiology by facilitating more predictive, preventive, and

participatory health care (Ghasemi et al., 2016; Hillman & Goldsmith, 2011; Jha &

Topol, 2016). The potential impact of AI in radiology must be better understood to be

adopted, implemented, and supported. Research will contribute to this process and

contribute to positive social change by introducing decision support solutions capable of

improving diagnostic accuracy and reducing health care costs.

My primary goal for this qualitative exploratory research study was to identify the

potential impact of AI on spine imaging interpretation and diagnosis. Chapter 1 includes

the fundamental elements of the study. These include the background, problem statement,

and research questions. The chapter also includes an introduction to relevant terms and

definitions to enhance the reader’s understanding of the research topic. I present the

theoretical perspectives and conceptual framework used for the design of the study. I also

discuss relevant assumptions, delimitations, and limitations in the study. Chapter 1 ends

with a discussion of the social significance of the research study along with a transitional

summary to the next chapter.

Background

The success of diagnostic imaging is highly dependent on technology and

protocol. Technological advances in radiology result in a perpetual cycle of discovery,

disruption, opportunity, and adaptation. Current innovations are rapidly transforming

radiology from a qualitative to a quantitative science with a growing capacity to obtain

molecular and physiologic measures (Jha & Topol, 2016; Quer et al., 2017). This

evolutionary process has led to unprecedented growth in the volume of multidimensional

data which has and will continue to have an impact on radiology workflow and the type

of decisions which have to be made. Exposure to an increasing quantity of complex data

increases the likelihood of diagnostic errors (Obermeyer & Ezekierl, 2016; Ragupathi &

Ragupathi, 2014). The growing demand to read imaging studies faster also influences

interpretive accuracy. For example, the average emergency radiologist may read up to

50,000 to 100,000 images per day, resulting in an average of 2 to 3 seconds spent on each

image (Syeda-Mahmood, 2015). Sokolovskaya et al. (2015) reported that radiologists

pushed to read complex diagnostic imaging studies at a faster speed made over twice as

many interpretive errors as those who read the same studies at a normal rate. Emerging

health care system productivity goals often drive the radiologist to interpret imaging

studies in a shorter period of time. The time constraints combined with the growing

complexity of studies adds to the burden of decision making.

The error rate in radiology has been a concern for decades. Pioneering research

performed by Garland (1949) over 60 years ago revealed an estimated 30% level of

inaccuracy regarding the interpretation of abnormal chest radiographs. This landmark

study prompted follow-up research. The error rate of radiologists on the interpretation of

abnormal imaging studies between 1949 and 1992 was estimated at 30%, consistent with

Garland’s work decades prior (Berlin, 1994; Renfrew et al., 1992). Additional studies

have revealed error rates for the interpretation of different types of abnormal imaging

studies in the range of 15–35% (Elmore, et al., 1994; Lehr et al., 1976; Janjul et al.,

1998). Interpretive errors in the current health care environment are potentially more

influential than they have been in the past because of widely distributed results within

electronic medical record (EMR) systems. Interpretive errors embedded within EMR can

lead to exposure at multiple levels along the health care chain, setting the stage for

additional errors in subsequent testing, clinical application, and judgment.

Paralleling the burden of big data in radiology there is a growing demand for

more precise imaging interpretation and concise reporting to support personalized care.

To successfully meet this demand, new decision support solutions have to be embedded

within the interpretive stage of radiology workflow. Success will require the integration

of human and machine attributes, a form of collective intelligence (CI). Radiomics, a

rapidly emerging AI supported process, is capable of performing high-throughput

automated analysis of imaging data through a series of sequential steps such as pathology

detection, feature extraction, feature characterization, and analysis (Lambin et al., 2017;

Lee et al., 2017; Limkin et al., 2017). Pathology features include physiological,

molecular, morphological, statistical, and textual attributes (Aerts, 2017). Successful use

of radiomics has the potential to help detect, characterize, and classify disease. It is

capable of revealing new signatures of pathology not available through visual

interpretation and contributing to probability-based decision support.

IBM developed a cognitive computing system in 2007 referred to as Watson,

which introduced to the public in 2011 as a potential AI solution for decision support in

the presence of complex data and uncertainty (Hoyt, Snider, Thompson, & Mantravadi,

2016). Watson was designed to augment the role of the radiologist during the interpretive

stage of radiology workflow and during the differential diagnostic process (Brink et al.,

2017; Jha & Topol, 2016; Kharat & Singhal, 2017; Jiang et al., 2017; Ranschaert, 2016).

AI solutions such as IBM Watson have the potential to improve diagnostic accuracy and

timing through the use of automated pattern detection, evidence analysis, and probability-

weighted scoring (Doyle-Lindrud, 2015; Ferrucci, 2012; Kohn et al., 2014). Some AI

systems are capable of learning with exposure to structured and unstructured data. This

includes exposure to data from radiomic methods, prior imaging studies, electronic health

records, disease databases, computational disease models, and peer-reviewed literature.

Interpretive and diagnostic errors in radiology occur due to many conditions

including technological limitations, system inefficiencies, and human error (Berlin, 2013;

Thammasitboon, Thammasitboon, & Singhal, 2013). Human perceptual errors occur

more frequently than cognitive errors (Berlin, 2013; Bruno, Walker, & Abujudeh, 2015;

Donald & Barnard, 2012). The causes of perceptual errors include reader fatigue,

distractions, and the presence of hidden or subtle disease patterns overlooked with visual

analysis methods (Bruno et al., 2015; C. S. Lee et al., 2012). Potential solutions for

reducing errors include redesign of the interpretive stage of radiology workflow to

include AI solutions, better integration of radiology-pathology data, 3D viewing options,

and structured reporting (C. S. Lee et al., 2012). The use of AI has already demonstrated

that it can improve the accuracy and quality of some decisions in radiology and in other

areas of health care (Brink et al., 2017; Jiang et al., 2017). The potential role of AI in

radiology is relatively new, subsequently; research and development will be influenced

by needs analysis and meaningful use cases. Research on the use of AI in spine care, and

more specifically spine imaging is extremely limited, despite its potential for having a

favorable impact on one of the most common causes of pain and disability.

The volume of personalized data acquired with advanced diagnostic imaging is

growing as the result of more detailed tissue interrogation, expanded fields of view,

thinner slice thickness, and the use of multimodality and multiparametric methods (Aerts,

2017; Hillman & Goldsmith, 2011; Jha & Topol, 2016; Quer et al., 2017). This trend will

influence all areas of radiology including spine imaging. The increased volume of

accessible personalized data contributes to the complexity of decision making

surrounding individual patient care and to inconsistencies and variability in radiology

reports.

According to Lundstrom, Gilmore, and Ros (2017), the interpretive stage of

radiology workflow may serve as the primary hub for the convergence and analysis of

imaging data, pathology data, and genomic data. This would empower the radiologist as a

curator of integrated patient data and as a clinical consultant. New decision support

solutions accessible during image interpretation will have a significant impact on the

timing and precision of the diagnosis process and subsequently decisions at the point-of

care. Research has demonstrated that AI use offers “higher than clinician-grade accuracy”

in numerous fields such as in dermatology, ophthalmology, and radiology (Syeda-

Mahmood, 2018, p. 573). AI has the potential to support more accurate, efficient, and

timely decision making in radiology. Its role in spine imaging needs to be addressed. AI

applications could help overcome errors associated with human bias while democratizing

expert decision support. I designed this study to identify how AI could improve

personalized spine care and help direct further investigation on what is required to

achieve this goal.

Problem Statement

The unprecedented growth of data in radiology is driven by new imaging

technology and methods capable of performing whole body surveys and evaluating

tissues at multiple length scales such as anatomic, physiologic, cellular, and molecular

levels (Aerts, 2017; Hillman & Goldsmith, 2011, Murdoch & Detsky, 2013; Raghupathi

& Raghupathi, 2014). Radiology is rapidly evolving into a dynamic, whole systems

diagnostic discipline capable of interrogating multiple dimensions of pathology in vivo

(D. Y. Lee & Li, 2009). The growing complexity and volume of imaging and non-

imaging data available during the interpretive stage of radiology workflow is exceeding

the individual radiologist’s ability to make fully informed decisions (Thrall et al., 2016).

Radiologists are finding it increasingly difficult to determine what imaging findings are

clinically significant and meaningful. Big data in radiology has become a burden which

can be overcome to create new opportunities for patient care. The primary problem

addressed by this research was defining how using AI during the interpretive stage of

spine imaging could augment the role of the radiologist and improve diagnostic precision.

As previously stated, the rate of error on abnormal radiology studies has been

reported as high as 30%; retrospective review of diagnostic imaging studies yielding even

higher error rates (Berlin, 2007, 2013; Brady et al., 2012, Donald & Barnard, 2012). An

overwhelming amount of data contributes to missed patterns and relationships resulting

in diagnostic errors with deleterious consequences (Gupta et al., 2017; Saber-Tehrani et

al., 2013). The causes of error in radiology include the burden of complex data,

increasing workloads; limited time, physiological fatigue, and human bias (Brady et al.,

2012; Donald & Barnard, 2012; C. S. Lee et al., 2012). The high level of inconsistency

and variability in radiology workflow is influenced by what data are acquired, how the

data are analyzed, and how the results are reported (Aerts et al., 2013). Several authors

have suggested that without adequate technological assistance, the human interpretive

process within radiology workflow will become progressively more inaccurate,

inefficient, and untimely (Croskerry 2013; Manrai et al., 2014; Obermeyer & Emanuel,

2016; Ragupathi & Ragupathi, 2014). Few studies have addressed how to reduce

interpretive errors in radiology (Fitzgerald, 2005; Lee, 2017).

I was unable to identify any peer-reviewed articles in which the authors

specifically address how to reduce interpretive errors associated with big data in spine

imaging. I was also unable to locate peer-reviewed publications in which the authors

address the use of AI methods such as radiomics combined with text analysis to reduce

interpretive errors in spine imaging. The spine-related complaints represent one of the

most common causes of chronic pain and disability in the United States (Hurwitz et al.,

2018; Loney & Stratford, 1999; Ricca et al., 2006). The American College of Radiology

(2016) recently acknowledged the need for further research to help identify how AI could

be used to access meaningful data, enhance the interpretive phase of radiology workflow,

improve diagnostic accuracy, and reduce errors.

The use of decision support solutions such as AI during the interpretive stage of

spine imaging workflow could lead to more accurate detection, characterization, and

diagnosis of pathology. Despite heightened awareness of the need for new decision

support solutions and knowledge of the potential benefits of AI use in radiology adequate

research on the potential impact on spine imaging has not been performed. Qualitative

exploratory research is required to lay the foundation for further research surrounding the

use of AI and to identify how benefits can be achieved.

Purpose of the Study

There is a considerable amount of meaningful data and insight embedded within

medical images, undetectable through routine visual analysis, and therefore not

considered in the diagnostic process (Gillies et al., 2016; V. S. Lee, 2017). Such data

embedded within non-imaging records could be made available during the interpretive

stage of radiology workflow for correlation with imaging findings. Missed data or

misinterpretation of data may lead to an error in diagnosis. Overlooked data and data

patterns can also lead to missed opportunities in care. The primary purpose of this

research study was to explore the potential of AI to favorably impact diagnostic accuracy

and precision during the interpretive stage of spine imaging workflow. The approach

includes the potential role of natural language processing and radiomic methods to assist

in the detection, characterization, and monitoring of pathology.

Research Questions

The primary research question for this study was: What are the opinions of

experts regarding the potential use and impact of AI intelligence during the interpretive

stage of spine imaging workflow?

The subquestions for this study follow:

1. How could the use of AI-supported methods (auto detection, segmentation,

radiomics, natural language processing) during the interpretive stage of spine

imaging influence the differential diagnostic process?

2. How could the use of AI-supported methods during the interpretive stage of

spine imaging influence disease classification and staging?

3. What AI solutions could be used to create interpretive priority in spine

imaging?

4. What will the future of spine imaging interpretation workflow look like?

5. Could AI-supported solutions such as radiomics be used to interrogate spinal

tissue in vivo and eventually lead to a virtual (digital) biopsy?

6. What are some potential advantages of an in vivo virtual “digital” biopsy over

a traditional needle biopsy in spine care?

7. How could the use of AI-supported augmented reality (AR) or virtual reality

(VR) enhance the evaluation of pathology in spine imaging?

8. What are some potentially “meaningful use” applications of AI in spine

imaging?

9. Which construct of the technology acceptance model (TAM) will likely have

a greater impact on AI adoption during the interpretive stage of spine imaging:

perceived benefits or perceived ease-of-use?

10. Which characteristics of innovations proposed by the diffusion of innovation

theory (DOI) will likely have the greatest impact on AI adoption during the

interpretive stage of spine imaging: complexity, compatibility and

interoperability, observed effects or trialability?

Theoretical Perspectives and Conceptual Framework

Adopting and using new technology is influenced by many factors, including

awareness of technology’s potential, ease of use, and interoperability with existing

technology and workflow (Khan & Woosley, 2011; Oye, Iahad, & Abrahim, 2012). The

potential impact of new technology must be studied in-depth using a flexible and iterative

approach. A rigid theoretical framework would not allow for adequate exploration of this

topic of study. Exploratory research offers a descriptive and inductive approach which

supports discovery, and helps reveal potential advantages and disadvantages surrounding

the use of new technology (Creswell, 2009; Patton, 1990; Yin, 2014). Effective

qualitative research design requires a conceptual framework to help align theoretical

perspectives, the research purpose, and research strategies (Creswell, 2007; Marshall &

Rossman, 2011; Patton, 1990). Conceptual boundaries can be developed through the

integration of ontological, epistemological, methodological, and structural perspectives.

This research study was inductive, beginning with a set of assumptions and

research questions leading to data acquisition and analysis. This study would have been

restricted by the rigid application of an a priori theory or hypothesis. I did not use a

deductive approach. A qualitative exploratory approach supports the investigation of

possibilities within the framework of different contexts and realities (Creswell, 2013).

Because of the potential depth and breadth of this topic of study, it would have been

unfeasible to perform the exploratory study through a single theoretical lens or

worldview.

AI represents a complex technology associated with numerous processes;

therefore, its potential impact must be evaluated from a pluralistic perspective. No single

theory can be used to fully explore the potential role and impact of AI during the

interpretive stage of spine imaging workflow. I used an interpretivist-constructivist

epistemological view to address the topic of study. The interpretivist perspective helped

operationalize how and why questions surrounding the use of AI. A constructivist

perspective supported the development of a proposed model for using AI solutions during

the interpretive stage of spine imaging workflow. Consistent with the work of Creswell

(2013) and Patton (1990), I used theoretical perspectives to help develop the conceptual

framework of the study and to develop research questions which aligned with the

research methodology and purpose.

The diffusion of innovations theory (DOI) proposed by Rogers (1995) represents

one of the most widely accepted theories for exploring how the attributes of innovative

technology influence its adoption, use, and distribution (Kahn & Woosley, 2011).

Furthermore, DOI offers insight about the stages of adoption and adopter categories

(Dillon & Morris, 1996; Rogers, 2003). The characteristics of innovations that influence

adoption proposed by Rogers (2003) include relative advantage, interoperability with

existing systems, the potential for technology evolution, reinvention, complexity, and

trialability. Additional factors that influence adoption include awareness of benefits

(knowledge), knowledge of advantages and disadvantages (decision), method of use

(implementation), and clinical utility (confirmation). DOI includes five adopter

categories: innovators, early adopters, early majority, late majority, and laggards (Rogers,

2003). The literature review summarized in Chapter 2 includes information regarding the

use of theoretical constructs and related conceptual perspectives introduced in this

treatise. I designed this study to obtain consensus-based perspectives from published

white papers and insights from purposively selected innovators and early adopters in

focus group sessions.

The technology acceptance model (TAM) proposed by Venkatesh (2008) is used

to address behavior and perceptions that influence willingness to use new technology.

The two primary constructs of TAM are perceived usefulness (PU) and perceived ease-

of-use (PEOU). The synthesis of DOI and TAM constructs have been previously used to

frame exploratory research of computer and information technology (Carter & Belanger,

2005; Y. Lee, Hsieh, & Hsu, 2011; Legris, Ingram, & Colerette, 2003). DOI represents an

effective theory for addressing variables surrounding technology adoption, whereas TAM

offers insight about post-adoption use and support (Hameed & Arachchilage, 2016).

Select constructs from DOI were combined with constructs from TAM to guide the

development of research questions to be used in focus group sessions and to help develop

an initial a priori list of thematic coding categories for data analysis.

A well-developed conceptual framework provides boundaries that help direct the

acquisition, management, and analysis of data during research. A conceptual framework

can also be used to assess the impact of new technology on professional behavior and

workflow (Bogdan & Bilken, 1982; Patton, 1999). The conceptual framework of this

study supported a contextual inductive and iterative process for exploring the potential

impact of AI on the interpretive stage of spine imaging workflow. I used the conceptual

framework for this study to choose data analysis strategies and to help guide the literature

search summarized in Chapter 2. I focused on literature that revealed AI technology

characteristics and the goals of early adopters of AI in radiology. Research questions

were developed to address some of the perspectives offered by DOI and TAM constructs.

The chosen method of inquiry in this study required answers to how and why questions

surrounding the potential benefits and challenges associated with AI use during the

interpretive stage of spine imaging workflow.

Nature of the Study

I performed this qualitative exploratory case study with an in-depth, inductive

approach to acquire and analyze data to address the potential impact of AI on spine

imaging interpretation and diagnosis. Qualitative exploratory case study research offers

an effective method for acquiring contextual knowledge surrounding the use of new

technology (Creswell, 2013; Ponelis, 2015; Yin, 2014). The approach is also effective at

revealing themes surrounding the potential impact and role of new technology (Bradley,

Curry, & Devers, 2007; Patton, 1999). Qualitative exploratory research studies have been

successfully used in radiology (Sanberg et al., 2012). The approach in this case offered

the level of holistic investigation required to reveal previously unknown variables

surrounding the use of AI during the interpretive stage of radiology workflow.

A well-defined unit of study helps direct qualitative research. It also helps frame

and inform research strategies. The boundaries of a qualitative case study may consist of

a relationship, situation, process or a culture (Miles, Huberman, & Saldana, 2014). The

primary unit of study for this research is a process defined as the interpretive stage of

spine imaging workflow. This includes the workstation, display technology, embedded

AI solutions, available data, and the role of the radiologist. I chose the unit of study to

identify how and why AI solutions should be used during the interpretation of advanced

spine imaging. I also designed the study to evaluate the potential impact of AI on the

current state of the unit of analysis.

Qualitative exploratory research is capable of revealing contextual complexities

not adequately addressed by explanatory or quantitative research methods (Ponelis, 2015;

Yin, 1984). More specifically, qualitative research can be used to reveal themes and

patterns, not revealed with restrictive quantitative methods (Dubin, 1969; Patton, 2002;

Ryan & Bernard; 2003). This study required the acquisition of data from multiple sources

including expert documents, reflective journaling, and focus group sessions. The focus

groups comprised radiologists and AI experts who met study inclusion and exclusion

criteria. I used numerous methods to improve the trustworthiness of the study.

Definitions of Relevant Terms

I have listed the operational definitions of selected terms and phrases to help the

reader understand the research topic, findings, and conclusions.

Algorithm: An algorithm represents a mathematical formula or program, used as a

set of instructions for computational analysis (Kakhani et al., 2017).

Artificial intelligence (AI): Artificial intelligence refers to the use of technology to

identify patterns, solve problems, and make predictions using biological-like approaches

(Jiang et al., 2017).

Augmented intelligence (AI): Augmented intelligence refers to the use of

technology to enhance human capability and to provide decision support (Liew, 2018).

Big data: Big data refers to a large volume of complex data that is difficult to

analyze utilizing traditional methods (Farooki, Almeida, & Saltz, 2016).

Collective intelligence (CI): Collective intelligence refers to the combined use of

AI solutions and human intelligence to identify a pattern, solve a problem, perform a

task, or make a prediction.

Computer-aided detection (CADe): Computer-aided detection refers to the use of

a computer system to identify a pattern within structured data, unstructured data, or on a

set of diagnostic images (Hillman & Goldsmith, 2011).

Computer-aided diagnosis (CADx): Computer-aided diagnosis refers to the use of

a computer system and computational methods of analysis to provide a list of differential

diagnostic possibilities or a specific diagnosis (Hillman & Goldsmith, 2011).

Data science: Data science represents an interdisciplinary field that uses

computational methods to analyze data, to reveal patterns, study processes or to make

predictions (Aerts, 2017).

Deep learning (DL): Deep learning represents a subset of machine learning

capable of becoming more accurate and capable with exposure to data through the use of

multiple hidden processing layers (Tang et al., 2018).

Differential diagnostic process: A differential diagnostic process is a series of

interrelated steps that use probability-based logic or reasoning to differentiate a disease or

disorder from others that may have a similar presentation.

Ground truth: Ground truth refers to data assumed or proven to be true (Choy et

al., 2018)

Innovation: Innovation is defined as a new or improved method of performing

tasks or solving problems (George et al., 2005).

In vivo: In vivo refers to the evaluation of biological elements, processes or

systems within a living organism (Lambin et al., 2012; Rizzo et al., 2018).

Machine learning (ML): Machine learning represents a subset of AI that uses a

computer system and a set of algorithms to reveal patterns in data with or without

handcrafted explicit instructions (Giger, 2018).

Natural language processing (NLP): Natural language processing refers to the use

of automated computational methods to analyze and interpret spoken or written language

(J. Y. Chen et al., 2017).

Neural networks: The phrase neural networks refer to a sophisticated network of

layered interconnected data paths and nodes, which process signals in a manner similar to

neurons in a biological system (Tang et al., 2018).

Precision medicine (PM): Precision medicine refers to an objective approach to

the diagnosis and delivery of health care which takes into account the heterogeneity of

disease along with an individual’s unique biology, disease risk, and variable response to

treatment (Mesko, 2017).

Radiomics: Radiomics refers to the science of high-throughput data analysis used

for identifying, extracting, characterizing, quantifying, and classifying nonvisible

elements of pathology within medical imaging data sets (Napel & Giger, 2015).

Segmentation: The definition of the term segmentation in radiology refers to the

use of manual, semiautomated or automated methods to identify and outline a region of

interest or an abnormality on images (Gillies et al., 2016).

Systems medicine: Systems medicine also referred to as network medicine refers

to the science associated with the relationships and interconnections between molecular

structures, cells tissues, and organs (McCue & McCue, 2017).

Virtual biopsy: The use of the phrase virtual biopsy in radiology refers to the use

of various computational and radiomic methods to reveal, extract, and quantify in vivo

characteristics of pathology derived from imaging data guided by a defined region of

interest (Echgaray et al., 2016; Lambin et al., 2012; Thrall, 2016).

Assumptions

Assumptions that are true or likely to be true represent beliefs or perspectives.

Assumptions influence research designs, methods, and conclusions. It is therefore

important to acknowledge assumptions that may influence this research process. I had to

consider the potential influence of numerous assumptions in this research study. I

assumed that by choosing consensus-based white papers (research documents) published

by reputable radiology organizations that I would acquire expert insights regarding the

potential role of AI in radiology now and in the future. I assumed that four to six

purposively selected participants (experts) placed into one of two homogenous focus

group sessions would be enough to reach topic saturation. In addition, I assumed that the

research participants would have sufficient knowledge and experience to answer focus

group questions and productively contribute to focus group discussions. I also assumed

that the research participants would adequately represent the levels of expertise held by

others in their relevant fields.

I assumed based on the inclusion and exclusion criteria of this study that the

participants would have a comfortable understanding of the research topic and goals. My

use of a purposive sampling method increased the likelihood that each research

participant had the experience and knowledge required to address the research topic

questions. Each of the participants works in the health care field, thus improving the

likelihood that they are aware of the importance of an honest and ethical approach to

research.

I assumed that participating radiologists had limited knowledge of the potential

role of AI in spine care and that AI experts had limited knowledge of the potential impact

of AI on spine imaging interpretation. The presumed knowledge and limitations of AI

specialists and radiologists opened the door for creative discussions and discovery in each

session. The level of dedication required of individuals within each group to reach expert

status infers a high level of passion and dedication to the subject; therefore, I assumed

that each of the participants could contribute valuable insight and opinions during the

focus group sessions. I assumed that, given the preparatory steps taken prior to data

acquisition, each of the research participants fully understood their rights and

responsibilities in the research project. The research assumptions helped guide me during

data collection from focus group sessions and from other sources.

Scope and Delimitations

The topic of delimitations refers to anticipated constraints that may arise during

the research design or while conducting research. Several delimitations arose in this case.

Research participants voluntarily participated in the focus group sessions; therefore, it

was likely that they were interested in the role of AI in radiology and spine care. The

defined unit of study and boundaries surrounding this study served as relative constraints

to data acquisition and interpretation. Participant inclusion and exclusion criteria also

represented delimitations. The moderator guide and questions developed for the focus

group sessions served as adaptable delimiters in the study.

The primary purpose of this research study was to explore the potential impacts of

AI use during the interpretive stage of spine imaging workflow. I did not investigate the

impact of AI on interprofessional relationships within the radiology department or its

influence on economic issues surrounding the use of AI in spine care. I also did not

address how AI might specifically impact radiologists of different backgrounds, training,

and skill levels. I did not design the research project to address the administrative

challenges or costs associated with developing, implementing, or supporting AI solutions.

Limitations

Research limitations refer to potential barriers or boundaries that investigators

cannot control. Limitations in this study included the available knowledge and expertise

from a small study population. Furthermore, the survey sample used in this study was

purposive and small, resulting in a data analysis process which is more descriptive than

inferential. The focus group participants’ levels of expertise influenced research study

conclusions. Limitations associated with lack of familiarity with the research topic were

relatively low given the highly qualified nature of the experts who participated. The

biases associated with the use of a single examiner posed a potential limitation. I reduced

this risk with the use of validation methods such as member checking, within and

between group analyses, reflective journaling, and triangulation of qualitative data

acquired from numerous sources.

This research study consisted of a small purposively chosen population of experts;

thereby, limiting the ability to generalize results. Determining whether sample size was

adequate was influenced by many factors such as the outcome of the research validation

methods. The type of methods used such as triangulation of data, member checking, and

peer review must be carefully considered when determining sample size (Patton, 2002).

The use of two research participant categories and two focus group sessions limited the

scope of this portion of the study. The potential role of AI in spine imaging is broad;

therefore, a single case study cannot address all relevant dimensions of the topic. In

addition, there was limited time and resources that restricted the duration of this study,

rendering it difficult to study the perceived benefits of AI use during the interpretive

stage of spine imaging over time.

Significance of the Study

Medical errors represent the third most common cause of death in the United

States (Makary & Daniel, 2016) and one of the most costly and avoidable health care

expenses (Saber-Tehrani et al., 2013). Diagnostic errors represent one of the most

common reasons for malpractice claims (Andel et al., 2012). Diagnostic imaging

represents a widely used method for detecting, characterizing, and monitoring disease.

Subsequently, many of the important decisions made in health care arise from diagnostic

imaging studies (D. Y. Lee & Li, 2009). The critical role and relevance of diagnostic

imaging in health care is increasing. This study addresses a gap in the research literature

associated with the potential impacts of AI on spine imaging workflow and the diagnostic

process. Using AI during the interpretive stage of spine imaging workflow could help

reduce human errors and favorably contribute to a more accurate diagnostic and reporting

process, leading to improved patient care. Research in other health care and radiologic

specialties has demonstrated that using AI can contribute more consistent, quantitative,

and actionable information to the final report, thus influencing point of care decisions

(Augimeri et al., 2016; Boone et al., 2015; Mohebian et al., 2017). AI solutions combined

with effective data governance and data management in spine imaging could result in

fewer interpretive errors and improved diagnostic precision.

Traditionally, radiology has relied upon the visual perception and interpretation of

the radiologist (Pinto & Brunese, 2010). Perceptual factors surrounding visual

interpretation represent a common cause of error. Factors that contribute to perceptual

errors include reader fatigue, distractions, the presence of hidden or subtle disease

patterns, and various form of human bias (Bruno et al., 2015; C. S. Lee et al., 2012). A

specialized application of AI, referred to as radiomics, has been successfully used by

radiologists to reveal the heterogeneity and in vivo features of pathology not detectable

by the radiologist during the visual inspection of a study (Aerts, 2017; et al., 2016; H. W.

Wu et al., 2012; Yip & Aerts, 2016). Radiomic methods improve the ability to classify

disease and stratify treatment approaches, all leading to more precise and personalized

patient care. In addition, AI solutions have the potential to improve the diagnostic

imaging process through the ability to detect subvisual pathology and to correlate data

from other imaging and non-imaging sources (Dreyer & Geis, 2017). This includes data

from electronic health records, published literature, genetic databases, as well as from

prognostic and predictive disease models. Digitization of in vivo and in vitro pathology

supports image sharing and access to remote data analysis and expert consultations.

Sharing of digital information also allows for consensus-based decision support.

Advances in medical technology have historically contributed to improved

methods of detecting and treating disease (Clinton, 2000). Integrated co-evolution of AI

supported methods such as radiomics will contribute to the discovery of new molecular

signatures and biomarkers of disease. This will lead to new standards for evidence-based

care in all fields including spine care. Successful integration of human and machine

intelligence during the interpretive stage of spine imaging workflow has the potential to

help control unsustainable health care costs, and support personalized care (Hillman &

Goldsmith, 2011; Jha & Topol, 2016; Kressel, 2017; Lee, 2017). In addition, AI is

capable of democratizing decision support that will aid underserved facilities,

underserved regions, and inexperienced radiologists.

New levels of expectations and knowledge surrounding the use of AI will

influence standards of care, which will ultimately benefit individuals, families, and

society. Greater use of AI-supported diagnostic methods such as radiomics will expand

classifications of disease and improve personalized care. The results of this research

study contribute to positive social change by identifying how AI use during the

interpretive stage of radiology workflow could favorably shape the future spine care.

Successful use of AI during spine imaging interpretation could result in early detection,

early intervention, better treatment outcome, and reduced direct, as well as indirect costs

associated with chronic pain and disability. Moreover, AI could facilitate collaboration

between spine care providers of all disciplines by exposing the fundamental basis for

disease, democratizing expert decision support, and by providing evidence-based

measures of treatment outcome. Widespread use of AI decision support will help

overcome some of the barriers to collaboration associated with human ignorance and

biases. I designed this study to reveal potential applications of AI in spine imaging, as

well as in other fields and specialties. This study introduces the concept of the digital

(virtual) biopsy that could have a profound influence on further investigation of this

concept and its application in all fields of health care.

Summary

Given the variability in people and their spine disorders, spine care delivery needs

to be more precise and personalized. Better integration of machine and human

capabilities during the interpretive stage of radiology workflow will contribute to earlier

detection of pathology, better characterization of pathology, and a more precise and

timely diagnosis. AI and related solutions used in other fields such as oncology and

cardiology can be adapted and used during the interpretive stage of spine imaging. The

primary goal of diagnostic imaging in spine care is to provide insight and knowledge that

can be used by a provider to deliver care for the right patient, for the right condition, at

the right time.

Imaging represents one of the most commonly performed and revealing elements

of the diagnostic workup in spine care. The combined burden associated with the growing

volume of imaging and non-imaging data available during the interpretive stage of

radiology workflow is increasing the complexity of decision making for the radiologist.

The missed opportunity and error rate associated with the interpretation of abnormal

imaging studies is too high. The growing complexity of data acquired through more

advanced imaging technology will only contribute to more complex decisions and higher

incidence of error. Applying AI solutions during the interpretive stage of spine imaging

workflow has the potential to detect early-stage pathology, offer decision support, and

facilitate personalized care.

The primary purpose of this research study was to identify whether the use of AI

solutions during the interpretive stage of spine imaging could reduce the risk for

diagnostic error and improve the precision of the final diagnosis. The potential AI

solutions introduced in this chapter and investigated in the next chapter include natural

language processing, radiomics, disease modeling, and computational analysis of

structured and unstructured data. The success of the interpretive stage of radiology

workflow is dependent on integrated solutions. It is becoming increasingly important to

use quantitative measures in diagnostic imaging. The integration of various decision

support solutions will likely change the landscape of radiology and spine care.

An accurate and precise diagnosis is dependent on heightened awareness of

possibilities and the ability to investigate the possibilities by analyzing available data.

Too often knowledge of differential diagnostic possibilities is limited by human

experience and limited access to technologies required to identify subtle or hidden

patterns within medical records or in the imaging data. This leads to errors in diagnosis

and clinical judgment. AI technologies such as natural language processing and radiomics

offer potential solutions. In Chapter 2, I address how current applications and

contributions of AI in radiology could be adapted or further developed for use during the

interpretive stage of spine imaging.

In this chapter, I introduced the research problem, research purpose, and research

significance, to help guide my literature search and review addressed in Chapter 2. The

subsequent chapter acknowledges current and potential applications of AI in radiology. In

Chapter 2, I synthesized the research purpose and goals from this chapter with the results

of an extensive literature search. This helped inform the research design and

methodology introduced in Chapter 3.

Chapter 2: Literature Review

Introduction

Exposure to large volumes of complex data in radiology increases the level of

uncertainty and complex decision making. This challenge combined with human

limitations and bias increases the risk for errors and oversights. The field of radiology has

always been dependent on the use of technology and has long served as a pillar in health

care for the acquisition, analysis, and management of complex data (Thrall et al., 2016).

Imaging has become one of the most important sources of data and diagnostic

information in health care (Aerts et al., 2013; Gillies et al., 2016: Kim et al., 2013;

Kinahan, & Hricak, 2016). The radiologist is poised to become a gatekeeper of big data

and a disease consultant (di Piro et al., 2017). This role will be augmented with the

convergence and integration of imaging, laboratory, genetic, and pathology data at the

radiology workstation.

The field of radiology is rapidly transforming from a predominantly qualitative to

a quantitative science supporting the rising demand for a more personalized diagnoses

(Jha & Topol, 2016; Quer et al., 2017). In addition to the demand for a more precise

diagnostic process, the radiologist is burdened with a growing quantity of complex data

arising from advances in whole body imaging, molecular diagnostic methods, and the

integration of multimodality and multiparametric imaging approaches (Murdoch &

Detsky, 2013; Raghupathi & Raghupathi, 2014). New forms of decision support are

required during radiology workflow to reduce the potential for interpretive and diagnostic

errors. The results of an extensive literature search revealed how AI solutions have begun

to have a significant impact on disease detection, characterization, and surveillance in

health care.

In this chapter, I discuss the literature search strategy that I used to address the

topic of study. I also address the fundamental concepts and research findings that I

identified in the literature that support qualitative exploration of the potential role of AI

use during the interpretative stage of spine imaging. Spine and spine-related research

involving the use of computational decision support such as AI were limited in number

and scope. The majority of published research and reviews addressed the role of AI

solutions in other fields such as oncology and neuroimaging. The literature review

identified prior methods of inquiry and research used to perform the studies. It also

revealed gaps in the literature regarding the use of AI during the interpretive stage of

non-spinal and spinal imaging. Scholarly publications provided the rationale for refining

the research problem, as well as developing the research questions and methodology for

this study. Given the limited number of publications referencing the use of AI in spine

imaging, the literature search was expanded to include research on the use of AI during

the interpretive stage of imaging other regions of the body. I chose and organized the

topics of this chapter to present relevant findings from the literature. The literature search

established a scholarly foundation for the research design and methodology covered in

Chapter 3.

Literature Search Strategy

I performed an extensive literature search to explore existing perspectives and

facts surrounding the topic of study. My helped to identify relevant theories, conceptual

frameworks, and research strategies that were applied in this study. The literature search

revealed current expectations and standards associated with the use of AI in radiology.

My literature search concentrated on articles that were published within 5 years of

the anticipated completion date of this work. My search was primarily limited to peer-

reviewed scholarly publications. I retrieved journal articles from the following research

databases; PubMed, EBSCO, ProQuest, and Google Scholar. I used key terms and

phrases along with different techniques to perform and refine the search process.

Common search terms and phrases included artificial intelligence, deep learning,

machine learning, radiomics, natural language processing, interpretive radiology, spine,

spine imaging, virtual biopsy, in vivo, voxel-wise detection, computer-aided detection,

computer-aided diagnosis, and biometric analysis. I used independent and combined

search terms to improve the investigative process.

The literature search was highly iterative to achieve adequate saturation of the

topic. I used single and combined Boolean operators, truncation, and wild card symbols

in the search process. I reviewed publication abstracts to ascertain article relevance. All

relevant research and review articles were printed, read, and saved. I evaluated the

bibliographic reference lists of seminal articles for additional works. My literature search

identified research studies, as well as consensus-based documents and position papers

published by respected institutions and organizations. During the literature review

process I was able to identify key experts on different research topics. I performed

author-based searches to look for relevant material they may have authored or contributed

to. Research librarians provided access to published articles I was unable to find through

more traditional methods. Thus, the literature search on the topic was exhaustive.

Applied Theoretical and Conceptual Framework

Successful adoption and use of new technology are influenced by knowledge of

its utility and its role within existing workflow (Khan & Woosley, 2011; Oye, Iahad, &

Abrahim, 2012). This knowledge is acquired through reading published works, listening

to colleagues or through hands-on experience. A rigid theoretical framework cannot be

successfully used to study the role of AI in radiology due to its numerous components,

rapid evolution and the complexity of its impact on the spectrum of workflow. As

previously stated, AI represents a complex technology associated with numerous

processes; therefore, its potential impact must be evaluated from a pluralistic perspective.

Exploration using a conceptual framework developed from the synthesis of constructs

and perspectives from different theories and models supported an adaptive and iterative

approach to addressing the potential role of AI in this study. This approach is consistent

with the work of Creswell (2009), Patton (1990), and Yin (2014).

The synthesis of DOI and TAM constructs have been used in numerous research

studies to explore the potential applications and benefits of new technology (Carter &

Belanger, 2005; Y. Lee et al, 2011; Legris, Ingram, & Colerette, 2003). In this case, the

synthesis of constructs from DOI and TAM led to the use of practical descriptive

categories such as perceived usefulness, perceived ease-of-use, clinical utility, diagnostic

accuracy, interoperability, and workflow compatibility. Each of these perspectives is

applied to the role of AI in radiology and is addressed through a variety of headings in

this chapter. In summary, my use of DOI and TAM perspectives supported the

development of a guiding conceptual framework used to help direct the literature review,

create focus group questions, and analyze acquired data. These conceptual perspectives

were also used to help frame the results of the literature search presented in this chapter.

The DOI proposed by Rogers (1995) acknowledges various categories of

technology adopters. These categories include the innovator, early adopter, early

majority, late majority, and laggards. Scholarly investigation of publications by

innovators and early adopters of AI use in non-spine related specialty fields of radiology

provided some of the insights required to address the topic of this research study.

Technology-Induced Transitions

The field of radiology has and will continue to be shaped by technology

development and its evolution (Hillman & Goldsmith, 2011). Advances in imaging

technology influence the timing and type of decisions, which have to be made during the

interpretive stage of radiology workflow. Innovations also restructure the realm of

expectations surrounding the analysis and flow of imaging data. Historically,

transformative technological advances in radiology have included radiology information

systems (RIS), picture archiving communication systems (PACs), natural language

processing (NLP), and more recently AI solutions (Hillman & Goldsmith, 2011). Each of

these technologies has improved some aspect of radiology workflow including the

detection, characterization, and reporting of disease (Sardanelli, 2017). The rapid

evolution of technology and data management in radiology is supported by the

miniaturization of materials, increased computer processing power, improved data

storage, enhanced connectivity, as well as greater access to disease models and registries.

Technological advances that offer decision support such as AI has and will

continue to influence every stage of radiology workflow. Computational decision support

influences the roles and responsibilities of radiologists. Widespread adoption and use of

AI-based decision support will require efficient integration with existing workflow and

legacy systems (Bauer, 2017; Dreyer & Geis, 2017). Learning how AI might improve

spine imaging interpretation initially requires exploratory research.

Computer systems and related AI solutions are often associated with numerous

components and processes, each of which contributes to new insights and technological

advances. This process is often referred to as co-evolution. Arthur (2009) introduced the

supporting premise that “existing technologies beget further technologies” (p. 21). The

impact of new technologies and their spinoffs is often underestimated, whereas the speed

of implementation is often overestimated. Knowledge and appreciation for AI technology

co-evolution is required to predict its impact on interpretive workflow.

Technological advances in radiology lead to perpetual cycles of discovery,

disruption, and adaptation. This unstable process introduces threats to established

protocols and standards (Lai, 2017). Each time new technology emerges, its potential

impact on the delivery of care has to be evaluated, and its clinical utility determined

(Kressel, 2017). In addition, the benefits and risks associated with an innovation have to

be disclosed and discussed. The attributes and influences of new technology such as AI

must be considered when evaluating its potential impact during the interpretive stage of

spine imaging workflow.

Potential AI solutions in radiology cannot be studied in isolation. They must be

studied in the context of other technologies and processes embedded into workflow. This

includes assessing its potential impact on diagnostic accuracy and efficiency. Zhang et al.

(2004) introduced a hierarchical system of human and technological relationships that

can contribute to or amplify medical error. The hierarchy includes the “individual,

individual-technology interaction, distributed systems, organizational structures,

institutional functions and overarching national regulations” (Zhang et al., 2004, p. 194).

The impact of new technology and decision support can have both favorable and

unfavorable outcomes. For this reason, the potential advantages and disadvantages of AI

use during the interpretive stage of radiology workflow must be considered within the

context of hierarchical professional and technological relationships.

Big Data Attributes and Related Burdens in Radiology

Diagnostic images are more than pictures. They contain massive quantities of

minable and potentially meaningful data, often not considered during the interpretive

process. Medical imaging is estimated to represent as much as 90% of all stored medical

data, contained within billions of images (Lambin et al., 2017). Imaging data is present in

many forms including symbols, words, images, and binary digits. Individual datum and

aggregations of data have unique characteristics that influence how it is acquired,

analyzed, and applied. The primary attributes of datum and data include value,

variability, veracity, velocity, and volume (Gandomi & Haiderm, 2015; Nendaz &

Perrier, 2012). The primary sources of data available during the interpretive stage of

radiology workflow are the study requisition, electronic medical records, diagnostic

imaging, and disease registries (Brown, 2014). In the near future there will be greater

access to relevant pathomic and genomic data at the radiology workstation. Human

interpretation of diagnostic images is often limited to visual analysis and qualitative

descriptive reporting (Thrall et al., 2016). Additional methods are required to analyze

textual and nonvisible data in radiology.

Data are commonly classified as structured or unstructured. Structured data have

well-defined form and context, whereas unstructured data tend to have inconsistent form,

rendering it more difficult to analyze. The most common type of unstructured data is

language, whether printed or verbal (Raghupathi & Ragupathi, 2014). The majority of

personalized data in health care are unstructured, in the form of text, represented in

medical records and reports.

Emerging computational methods are being used in radiology to help transform

data and related patterns to meaningful information or knowledge that has clinical utility.

Determining data relevancy in radiology requires knowledge of its validity and clinical

utility. The majority of acquired data in radiology are considered noise and are not

relevant to clinical care. This perspective is influenced by current limitations in pattern

detection and limited human capacity to analyze nonvisual imaging data (Kohn et al.,

2014). The large volume of data acquired during current imaging studies only represents

a fraction of what will be acquired, mined, and transformed into action in the near future

(Kohn et al., 2014). What are considered irrelevant and meaningless data now may

represent actionable data in the future.

A considerable amount of data and insight are embedded within medical images,

but remain undetectable through traditional visual analysis (Gillies et al., 2016; Lee et al.,

2017). As a result, radiologists face immense challenges during the interpretive stage of

radiology workflow. Despite these current challenges advances in diagnostic imaging

continues to create progressively larger and more complex data sets. Without adequate

technological assistance human interpretation of complex imaging data will become

progressively more inefficient, inaccurate, and untimely (Crosskery 2013; Manrai et al.,

2014; Murdoch & Detsky, 2013; Obermeyer & Ezekierl, 2016; Ragupathi & Ragupathi,

2014; Weber et al., 2017). As previously stated, new solutions are required. In the future,

whoever has access to the best data and best interpretive solutions will likely provide the

best care.

Decision-Making Processes

Clinical decision making in health care including radiology is often challenging,

and associated with complex nonlinear information (Hussain & Oestreicher, 2017).

Available methods to simplify the process include the use of published guidelines,

professional collaboration, crowd-sourcing, and computational decision support (Nendaz

& Perrier, 2012; Phua & Tan, 2013). Many health care providers, including radiologists,

are often confronted with decisions they are unprepared or unqualified to make.

Radiologists like other health care providers tend to look for what they know, identify

what they are familiar with, and render decisions based on experience. Variables such as

the quantity and quality of data, experience, professional knowledge, incentives, and

access to technological support influence human decisions (Gandomi & Haider; Weber &

El-Kareh, 2017; H. W. Wu et al., 2012). Accurate and timely decisions require awareness

of a well-defined problem, knowledge of options and alternatives, and the capacity to

evaluate the problem (Croskerry, 2013; Kohn et al., 2014). In addition to addressing

diagnostic variables, a radiologist’s decisions must meet the standard of care and be

consistent with a patient’s needs, values, and expectations. The diagnosis offered by a

radiologist after interpreting a set of images should be delivered in a manner that provides

adequate decision support for the referring health care provider at the point of care.

Decision Support Solutions

Decision support refers to a process or technique used to help determine the right

course or courses of action. The primary purpose of clinical decision support (CDS) in

radiology is to avoid errors, improve diagnostic accuracy, and improve the quality of care

delivered to the patient. A successful decision support system requires various attributes

such as availability, ease-of-use, accuracy, and consistency (Kahn, 1994). Decisions

made during image interpretation, the reporting process, and at the-point-of care

(Raghupathi & Raghupathi, 2014). The rate of discovery and knowledge creation

generally outpaces the individual health care provider’s ability to keep up to date and

make fully informed decisions (Kohn et al., 2014; Thrall et al., 2016). This phenomenon

applies to data intensive specialties such as radiology; therefore, the radiologist must

remain aware of decision support options.

Numerous solutions have been proposed to address complex decision making in

radiology. One of the more recent solutions is the computerized decision support system

(CDSS). CDSS solutions are placed into one of two categories; knowledge-based systems

or non-knowledge-based systems (Stivaros et al., 2010). Knowledge-based systems

contain programmed rules, an inference engine, and a well-defined communication

mechanism. Non-knowledge-based systems often use machine-learning techniques or

algorithms that learn from the ground up through the exposure, assimilation, and analysis

of available data. Effective use of decision support in radiology will help deliver the right

information, in the right format, at the right level of workflow to the right person. In

summary, access to decision support can augment the role of the radiologist.

Heuristics

Heuristics refers to the application of rules or processes to simplify decision

making. Radiologists often rely on heuristic methods such as mental shortcuts to

minimize delay, reduce task complexity, and to simplify decisions (Itri & Patel, 2018;

(Tversky & Kahneman, 1974). Heuristic methods are used to improve the accuracy, as

well as the efficiency of the interpretive process in radiology. Inappropriate use of

heuristics or the use of inaccurate heuristic methods can result in errors that adversely

influence patient care.

The three principal categories of heuristics are anchoring, availability, and

representative (Tversky & Kahneman, 1974). Anchoring heuristics refers to the limitation

of further considerations due to perceived truth. In contrast, availability heuristics refers

to the assignment of value based upon an individual’s memory or recall. This form of

heuristics can result in errors secondary to limited or selective memory of prior events or

outcomes. Representative heuristics is characterized by the use of categories to simplify

data, information, and knowledge. This approach may be used to simplify a process

leading to oversight and conjunction fallacy.

Dual Processes: Intuition and Analytics

The dual process theory of decision making consists of intuitive processing (type

I) and analytic (type II) approaches (Croskerry, Petrie, Reily, & Tait, 2014). An intuitive

response is characterized by a low cognitive demand and rapid application, whereas an

analytic approach is characterized by a high cognitive demand, a slow process, and

greater reliance on working memory. There are risks associated with isolated application

of one decision-making process over another (Phua & Tan, 2013). Croskerry (2014)

acknowledged the importance of discriminant use of both methods during complex

decision making. The combined use of problem-solving methods increases the likelihood

of a good decision and a good outcome. This perspective applies to human and machine-

based approaches.

Sources of Interpretive and Diagnostic Errors in Radiology

Diagnostic errors are often underreported and underappreciated due to a lack of

standards in defining, recognizing, and acknowledging their presence (C. S. Lee et al.,

2012). It is difficult to estimate the impact of radiology errors due to the reasons

mentioned and the limited capacity to measure their short and long-term impact on

overall health (C. S. Lee et al., 2012). Errors can occur anywhere along the path of

radiology workflow (Huassian & Oestreicher, 2017; Kassier & Kopeman, 1989). Errors

that occur during the interpretive stage of radiology workflow are likely to have the

greatest impact on the final diagnosis and report.

Health care providers too often make point of care decisions with irrelevant,

incomplete or incorrect information (Kelly & Hamm, 2013). In addition, many physicians

have limited experience with uncommon diseases or complex presentations associated

with coexistent pathology (Manrai et al., 2014). This condition can influence the

radiologist during image interpretation and can influence the referring physician who

reads the report at the point of care. According to Latts (2016). it is impossible for a

single health care provider to stay up-to-date in their field and to remain aware of all of

the relevant data and variables associated with any particular disease process or state.

This position supports the need for better decision support along the path of care.

Types of Errors

A medical error represents a deviation from a consensus opinion or a standard of

care. Errors may occur secondary to missed presentations, oversights, or mistakes of

judgment, all of which could lead to failure to implement a process or a plan of action

(Andel et al., 2012; Makary & Daniel, 2016). One of the most common forms of error in

radiology is diagnostic error, surfacing as a missed diagnosis, wrong diagnosis, or an

untimely diagnosis. Diagnostic errors in radiology have generally been classified as

perceptual or cognitive (Berlin, 1996, 2013). Several researchers have acknowledged

perceptual error as the most common cause of diagnostic errors in radiology (Berlin,

2013; Bruno et al., 2015). In support of this premise, a large radiographic research study

revealed that 80% of diagnostic errors were perceptual and 20% were interpretive

(Donald & Barnard, 2012). Causes of nondiagnostic errors in radiology include failure to

recommend or perform an indicated test or test protocol and failure to address clinical

concerns or reported patient presentations on imaging test requisitions.

Y. W. Kim and Mansfield (2014) proposed one of the most widely accepted

classification of errors in health care. The approach represented an expansion of

classifications previously proposed by Renfrew (1992), a few decades earlier. According

to Kim and Mansfield (2014) common causes of error in radiology include faulty

reasoning, lack of knowledge, satisfaction of search bias, miscommunication, and an

acquired inaccurate or incomplete history. Cognitive based-errors include limited skills,

stress, bias, faulty heuristics, memory loss, and inattention (Zhang et al., 2004). All types

of errors are amplified in the presence of large volumes of complex data. In addition to

the reasons given, diagnostic errors may occur as the result of technological limitations,

restricted access to data, and system failure (Berlin, 2013; Thammasitboon et al., 2013).

Errors in radiology may also occur secondary to cognitive bias rather than the result of a

perceptual error or lack of knowledge (Hussain & Oestreicher, 2017; Nendaz & Perrier,

2012). The cause of interpretive error in radiology is often the result of more than one

factor.

Expanding domain knowledge combined with human variables such as limited

time and preoccupation with the care of complex patients, contributes to a high incidence

of diagnostic errors within any specialty field (Weber & El-Kareh, 2017). Radiologists

are more likely to make wrong decisions in the presence of signs, symptoms or

conditions, which they have, limited experience with or knowledge of (Weber & El-

Kareh, 2017). Overwhelming data combined with limited knowledge and competing

environmental pressures contribute to chaos and confusion, which can lead to oversights

and diagnostic errors (Gupta et al., 2017; Saber-Tehrani et al., 2013). Access to

computational decision support such AI increases the potential for problem solving in the

presence of human limitations, as well as cultural and environmental pressures.

Human Limitations

Human decisions made in the presence of high degrees of complexity and

uncertainty increase the likelihood of error (Stivaros et al., 2010). The ability to process

information is limited by physiological mental capacity, a phenomena often referred to as

the cognitive threshold. Radiologists interpret images using a variety of cognitive

methods such as visual detection, pattern recognition, memory, and reasoning. A

radiologist’s cognitive performance is influenced by personal attributes, physiology, and

professional skills. It is also influenced by environmental factors such as ambient noise,

workload intensity, and workflow distractions. Stress influences interpretive accuracy.

Stress may be associated with uneven work distribution, poor reimbursement, limited

time, increasing liability pressures, along with heightened awareness of the complexity

and heterogeneity of pathology.

Humans are flawed in their capacity to process large volumes of multidimensional

or deeply nested data (Hatt et al., 2017; Wolf et al., 2015). The human brain is also

limited in its capacity to perform highly scalable functions that involve voluminous or

unrecognized confounding variables (Wolf et al., 2015; H. W. Wu et al., 2012). Humans

are limited in their capacity to perform accurate complex data analysis in a relatively

short period of time. Factors that contribute to these limitations include physiological

fatigue, cognitive bias, limited knowledge, and distractibility (H. W. Wu et al., 2012). A

conscious or unconscious response to self-limitations may result in the use of heuristic

methods that introduce bias to a decision-making process.

Increased complexity in a work environment increases the likelihood of human

error compounded by deficiencies of the system (Institute of Medicine, 2000). For this

reason, health care specialists responsible for analyzing large quantities of complex data

such as radiologists are often exposed to high cognitive demands and subsequently high

rates of diagnostic error (Crosskery 2013; Nendaz & Perrier, 2012; Obermeyer &

Ezekierl, 2016; Ragupathi & Ragupathi, 2014; Weber et al., 2017). Technological

solutions can be implemented to reduce and simplify the differential diagnostic process

by performing pre-analytic functions prior to human interpretation of images.

Inconsistency and Variability

Conventional imaging interpretation and reporting methods are highly subjective;

and subsequently, associated with a high degree of variability (Bosmans, Weyler,

DeSchepper, & Parizel, 2011; Bruno et al., 2015). Most radiology reports primarily

consist of subjective narrative descriptions of normal and abnormal findings (J. Y. Chen

et al., 2017). Moreover, radiologists vary in their use of interpretive descriptors and

reporting structures (Napel & Giger, 2015). This personalized approach to reporting

contributes to inconsistencies within and between radiology reports. In support of this

premise, radiology reports have been described by numerous researchers as incomplete,

inconsistent, and inconclusive (Bosmans et al., 2011; J. Y. Chen, Sippel-Schmidt, Carr, &

Kahn, 2017). In one particular example, variability of the interpretation of spine imaging

by different radiologists was reflected by a misinterpretation rate of 43.6% plus or -11.7

(Herzog et al., 2017). The study was based on the interpretation of 10 MRI studies

performed on the same patient at 10 MRI centers, read by 10 independent radiologists.

The degree of inconsistency and variability occurring during the interpretive stage of

radiology workflow in all fields including spine care must be improved.

In addition to variability of the reporting process, there is also a high level of data

management and data access variability during interpretive workflow (Aerts et. al.,

2013). Factors, which adversely influence the flow of data and the pattern of data access

during radiology workflow, include incomplete access to medical records and prior

imaging reports, inadequate imaging protocols, human error, and technical workstation

deficiencies. Image interpretation and the description of pathology often vary between

radiologists. The degree of interpretive variability is influenced by a radiologist’s level of

experience, the time allowed for the interpretive process, the complexity of the study, and

the presence of human bias (Napel & Giger, 2015). Interpretive variability is also

influenced by reporting requirements. AI has the potential to improve patient care by

improving the flow of data and access to data across the spectrum of radiology workflow

(Augimeri et al., 2016; Boone et al., 2015). AI can also influence how radiology reports

are structured.

Cognitive Bias

Cognitive bias represents an error in reasoning. Over 100 different forms of

cognitive bias have been identified in health care (Croskerry, 2017). Common forms of

bias in radiology include availability bias, alliterative bias, anchoring bias, framing bias,

satisfaction of search bias, and pro-innovation bias. Availability bias occurs when a

decision is influenced by experiences, whereas alliterative bias occurs when an

individual’s judgment is influenced by another. Anchoring bias refers to limiting the

search for additional possibilities due to the belief that a prior assumption is correct or

that a current diagnosis fully explains a patients presentation (Tversky & Kahneman,

1974). Framing bias refers to the use of a limited perspective. Satisfaction of search bias

refers to the assumption that the diagnostic process is complete due to lack of knowledge

of other differential diagnostic possibilities. Pro-innovation bias refers to assigning value

to the role of new technology or the data it provides without consideration for potential

inaccuracies or inconsistencies (Bauman & Martigoni, 2012; Rogers, 2003). Radiologists

may not be aware of their own cognitive bias under different circumstances. Machine-

based decision support is not subject to most forms of human bias, and therefore can be

used to help avoid or overcome adverse consequences of human bias.

Research has revealed that radiologists, like experts in other fields, are subject to

a phenomenon referred to as inattentional blindness, characterized by missing what

should be obvious due to search bias (Drew et al., 2013; Memmert, 2006). For example,

Drew et al., (2013) revealed that 83% of radiologists asked to review computed

tomography lung scans for nodules and other abnormalities failed to identify a gorilla

image located in the lung field. The gorilla image was more than 40 times larger than the

average nodule. This acclaimed research study confirmed that a prioritized search for

specific pathology could lead to blinding of other significant findings.

Cognitive bias can influence data acquisition, data analysis, and data

interpretation during the course of radiology workflow. The primary factors that

influence cognitive bias include poor training, lack of experience, stress, uncertainty,

physiological fatigue, and incomplete information (H. W. Wu et al., 2012). Different

forms of bias may overlap or coexist within the same decision-making process. For

example, anchoring bias may be amplified by confirmation bias leading to premature

diagnostic closure (Hussain & Oestreicher, 2017). The use of detrimental heuristics may

complicate cognitive bias and result in higher risk for diagnostic error.

Process and Workflow Error

Any situation that disrupts or interrupts the interpretive process during radiology

workflow can lead to human distraction and diagnostic error. Examples include slow data

access, difficulty accessing prior imaging studies or records, and lack of familiarity with

complicated workstation technology. Additional distractions include phone calls,

interventional procedures, and conversations with health care providers (Schemmel et al.,

2016). Numerous disruptive factors commonly occur simultaneously or within a short

time frame in the radiology setting.

The absence of technological decision support in the presence of complex data

can result in an inefficient, inaccurate, and untimely diagnostic process. Potential

solutions include the use of embedded AI decision support, access to integrated radiology

and pathology data, physical workflow modification, and simplified workstation

interfaces (Bruno et al., 2015; C. S. Lee et al., 2012). It is important to embed

technological solutions within radiology workflow to simplify the interpretive process

and to reduce the risk for diagnostic errors.

Artificial Intelligence: An Introduction

Artificial intelligence (AI) refers to technology that exhibits biological-like

properties to assist, augment or replace human processes or actions (Pannu, 2015). The

Turing test developed decades ago offered an operational definition of AI, which required

that a machine possess certain human-like attributes and capabilities that could be used

for problem solving (Turing, 1950). The two principal forms of AI are machine learning

and natural language processing (Jiang et al., 2017). Natural language processing (NLP)

is used for textual analysis, whereas the use of machine learning (ML) in radiology

supports computer-aided disease detection, characterization, and monitoring. ML can be

used to assist in the differential diagnostic process.

The three primary forms of AI are assisted intelligence, augmented intelligence,

and autonomous intelligence (Bothum & Lancefield, 2017). Rapid advances in computer

technology, software programming, and algorithm development have accelerated the

evolution of AI in the direction of autonomy for some tasks. Assisted intelligence refers

to the use of technology to improve a process a human is capable of performing. In

contrast, augmented intelligence refers to the use of technology to enhance human

potential. Autonomous intelligence refers to the use of technology to perform a task that

exceeds human capability. One of the most important attributes of AI is speed. It can

perform most tasks much faster than humans can. Some AI solutions will evolve from

offering assistance to becoming autonomous. The primary elements of an intelligent

system include infrastructure, algorithms, data, software, and an ecosystem.

The broad topic of AI encompasses different computational methods such as

machine learning (ML), deep learning (DL), and cognitive computing (CC). The

numerous subcategories of AI, each of have different potential applications and response

characteristics (Figure 1). The basic elements of expert machine systems include a

knowledge base, an inference engine, and established rules operationalized by algorithms

(Salem, 2017). Algorithms represent digital rules used to perform an automated task or

operation. DL algorithms have many applications in radiology. For example, they can be

used to reveal new features of disease, not previously identified. Conventional ML

algorithms are linear, whereas DL and CC algorithms are more abstract, characterized by

a hierarchy of increasing complexity. Leading radiology companies have adopted

different terminologies for their AI solutions. For example, Phillips refers to its AI

solution as adaptive intelligence, General Electric refers to its solution as applied

intelligence, and IBM refers to its solution as cognitive computing (Freiherr, 2018).

Despite the use of different terms, the goal is to use technology to enhance or replace

human performance for improving a process and/or an outcome.

Traditional computer programming requires the use of explicit rules to perform

tasks. In contrast, ML uses statistical techniques and algorithms that do not require

explicit rules (Cai et al., 2016). The processes associated with ML performance can be

classified into three primary categories, which are supervised learning, unsupervised

learning, and semi-supervised learning. Supervised learning refers to the use of expert

derived handcrafted rules that serve as map for the flow of data between input, output,

and ground truth. Unsupervised learning refers to the use of one or more algorithms

designed to reveal patterns within data without a priori rules or human intervention.

Semi-supervised learning represents a combination of both approaches. Supervised

learning is often used to train a model to make a prediction, whereas unsupervised

learning methods are often used to explore data without a preconceived determination.

Fundamental ML data analysis methods include classification, regression, clustering,

pattern matching, density estimation, and dimensionality reduction (Kotsiantis, 2007).

ML systems can be used in radiology to detect patterns, aggregate data, and classify

disease features.

Figure 1. There are different forms and applications of artificial intelligence. Each

subtype in the above figure is less dependent on handcrafted rules and more capable of

detecting patterns in complex data and learning with exposure to data.

Deep learning (DL) represents a subset of machine learning, modeled after

neurological signal transmission in the nervous system (Zaharchuk et al., 2018). One of

the most common forms of DL is referred to as convolution neural networks (CNN),

characterized by an advanced network of connectivity capable of performing complex

parallel and serial processing (Cascianelli et al., 2017; Dreyer & Geis, 2017; Zaharchuk

et al., 2018). DL methods are able to detect patterns within high-dimensional data sets

using layered processes that impart logic (Lakhani & Sundaram, 2017). Neural networks

have generally outperformed individual algorithms in the analysis of complex data

(Lerner et al., 2018). DL solutions tend to become more accurate with exposure to data

due to their ability to detect patterns and create supportive algorithms (Zaharchuk et al.,

2018). DL methods have been successfully used in different health care fields including

neuroradiology. For example, Gao et al., (2017) used CNN for the automatic

classification of 285 non-contrast brain CT examinations into one of three categories of

pathology. The autonomous capacity of DL systems to learn and to construct pattern

detection algorithms renders them capable of detecting previously unrecognized patterns

within complex imaging datasets.

Algorithms represent a programmed set of rules delivered through a sequence of

operations to manipulate and analyze data to achieve a desired outcome (Obermeyer &

Emanuel, 2016). They may be handcrafted or created by a computer. Algorithms

essentially represent mathematical formulas that are used as instructions for a digital

process (Mapoka, Masebu, & Zuva, 2013). They are used to detect patterns, sample

variables, register instances, highlight structures, and create reference maps. Algorithms

may be deterministic, logical, or recursive. Their success is dependent upon their

relevance, accuracy, and speed. Algorithms render computer systems capable of

augmenting the role of the radiologist.

The most common types of algorithms used in radiology are k-nearest neighbors,

convolutional neural networks, fuzzy logic, support vector machines, decision trees, and

Naïve Bayes algorithms (Adduru et al., 2017; Erickson, Korfiatis, Akkus, & Kline, 2017;

Fan, Lin, & Tang, 2017). Algorithmic outcomes can be expressed in many forms, which

include statistics, natural language, flowcharts, diagrams, and a list of probability-based

differential diagnostic possibilities. The handcrafted algorithm sometimes serves as a

computational bridge between human and machine processes to transform complex data

into practical information (Beam & Kohane, 2017; Obermeyer & Ezekiel, 2016).

Algorithms can be used to sort through the millions of variables and patterns within

advanced imaging datasets. They can also be used to correlate structured and

unstructured data within radiology workflow.

AI uses two primary types of algorithms during computational analysis. These

algorithms are referred to as classifiers and controllers. Classifiers are used for pattern

detection and pattern matching, whereas controllers are used to assign an action or task to

computational outcomes. Special digital filters combined with algorithmic functions have

been used to mine and interrogate data (Obermeyer & Emanuel, 2016). Algorithms are

not subject to cognitive bias but have the potential to introduce machine bias during their

creation and evolution. Machine bias can lead to false negative or false positive

outcomes.

Introduction to IBM Watson: A Cognitive Computing System

IBM Watson is a cognitive computing system that was introduced in 2015 to offer

solutions for problem solving and decision making in the presence of complexity and

uncertainty (Hoyt, Snider, Thompson, & Mantravadi, 2016). The system is capable of

generating “ descriptive, predictive, and visual analytics, thus, reducing the risk for error”

(Hoyt et al., 2016, p. e165.). Cognitive computing is capable of analyzing structured and

unstructured data; thereby, increasing its potential utility in radiology.

IBM Watson represents the first large-scale integrated clinical diagnostic support

system (CDSS) capable of analyzing structured and unstructured data, and rendering a

differential diagnostic list based on pattern detection and probabilistic calculations. The

system uses a series of interdependent computational methods to respond to inquiries.

The process generally begins with a question followed by hypothesis development,

evidence analysis, and probability assignment (Deloitte, 2015; Ferrucci, 2012). With

further development, IBM Watson may become capable of augmenting the role of the

radiologist during image interpretation (Brink et al., 2017; Jha & Topol, 2016; Kharat &

Singhal, 2017; Jiang et al., 2017; Ranschaert, 2016). In summary, IBM Watson has the

potential to auto detect abnormalities, filter irrelevant or normal imaging data,

characterize disease, and assist in the differential diagnostic process with probability

assignment. Watson is capable of offering data and knowledge-driven decision support

that can be used to assist, augment or in some cases replace the role of the radiologist

(Kohn et al., 2014). Furthermore, IBM Watson is a pioneering solution capable of

integrating ML and NLP applications during the interpretive stage of radiology workflow

(Jiang et al., 2017). IBM technology simply serves as an example of what AI is capable

of achieving in health care. Many products will be developed to perform the same or

similar tasks. It is too early to tell which technologies will meet the reproducibility and

validation demands of future research and regulatory requirements.

Successful adoption and use of AI solutions such as IBM Watson in radiology

requires validation studies, heightened awareness of its potential value, and an adequate

state of readiness. AI solutions have the potential to reveal biological variability and the

features of pathology in vivo in a manner which exceeds human capabilities (Gillies et

al., 2016; Limkin et al., 2017; Pannu, 2015; Yip & Aerts, 2016). AI solutions also have

the potential to calculate and assign probability to differential diagnostic possibilities

based on analysis of structured and unstructured data. For these reasons, AI solutions

should be developed to improve the precision and personalization of the diagnostic

process.

Collective Intelligence: A Collaborative Approach

Data can be analyzed and health care decisions can be made using collective

intelligence (CI), a process referring to the integration of human expertise and computer

analytics. The capabilities of CI include reducing the complexity of data, revealing

patterns within the data, assigning value to classifications of data, and offering decision

support (Hoyt, Snider, Thompson, & Mantravadi, 2016). The future impact of human-

machine collaboration is difficult to predict because human performance tends to

improve in a linear and relatively predictable fashion, whereas the capabilities of AI grow

exponentially. Advanced AI systems can learn from mistakes and successes and are

therefore less likely to repeat mistakes than humans are. Humans add an element of

creativity and intuition to AI output.

Unlike computers, humans have a limited capacity to analyze complex data.

Humans also have difficulty correlating multidimensional nonlinear variables,

performing syntactic transformations, and revealing patterns within variable high velocity

data (Nendaz & Perrier, 2012). Humans are prone to physiologic limitations and fatigue,

whereas computational technology is stable and consistent (El-Kareh et al., 2013; Russel

& Norvig, 2010). Human intelligence is highly dependent on experience, analytic skills,

intuition, and motivations (El-Kareh et al., 2013). In contrast, AI systems are highly

dependent upon access to annotated training data, ground truth, and validation methods

(Russel & Norvig, 2010). Human intelligence has many characteristics not offered by AI

such as adaptability, intuition, creativity, flexibility, and the ability to plan.

AI systems are not capable of replacing the full breadth and depth of human

reasoning and judgment. For example, the human role in radiology offers the benefits of

unique experiential insight, empathy, and thoughtful decisions. The combination of

human and artificial intelligence offers a collaborative approach with a greater chance of

success then either isolated approach in some settings. A collaborative relationship

supports what the radiologist does best and combines it with what AI does best.

Radiology Workflow Defined

The radiology workflow environment is complex. Each stage of radiology

workflow has unique elements or processes (Figure 2). Radiology workflow has been

divided into two primary stages referred to as interpretive workflow and non-interpretive

workflow (Lee et al., 2017; Schemmel et al., (2016). I divided the non-interpretive stage

into two subsets referred to as the pre-interpretive and post-interpretive stages. AI is

poised to play a significant role within all stages of radiology workflow, although its

greatest potential is likely within the interpretive stage, the focus of this research study.

AI solutions can be used during the interpretive stage of radiology workflow to detect,

characterize, classify, and monitor disease. Reasoning during the diagnostic process will

involve various types of diagnostic inference and decision support. AI support during the

interpretive stage of radiology workflow can also be used to provide a probability-based

differential diagnostic list for the attending radiologist.

Figure 2. The three stages of radiology workflow that are pre-interpretive stage,

interpretive stage and post interpretive stage. The figure also highlights elements of the

interpretive stage of workflow, which can benefit from AI support.

Many radiologists spend a greater amount of time performing non-interpretive

tasks than interpretive tasks during the course of radiology workflow (Dhanoa et al.,

2013). This pattern of labor contributes to inefficient and occasionally inaccurate

interpretive outcomes. Limited time combined with exposure to large volumes of

complex data exposes the radiologist to a higher degree of uncertainty and subsequently

to greater potential for the influence of human bias and error during the interpretive stage

of radiology workflow. There are numerous forms of human bias, which can take place

during the interpretive stage of radiology workflow (Figure 3). Quite often more than one

form of bias will be present. AI-based decision support can reduce the impact of human

bias.

Non-interpretive human tasks during radiology workflow include setting image

protocols, supervising studies, directly caring for patients, accessing analytic tools,

performing image-guided intervention, and consulting with health care providers. Human

tasks typically performed during the interpretive and post-interpretive stages of radiology

workflow include visual evaluation of images and qualitative report generation. The use

of AI methods during the interpretive stage of radiology workflow could help reduce the

incidence and prevalence of interpretive and diagnostic errors (Busby, Coutier, &

Glastonbury, 2018; Itri & Patel, 2018). Radiomic methods are increasingly becoming a

more important quantitative measure in diagnostic imaging, one that may change the

landscape of health care. The success of AI use in radiology will be dependent on its

ease-of-use, utility, and ability to be integrated into existing workflow.

Figure 3. Types of human bias that can occur during the interpretive stage of radiology

workflow.

Emerging Options for Interpretive Radiology Workflow

The current tasks performed during the interpretive stage of radiology workflow

must be adapted or replaced to support a more accurate and timely diagnostic process in

the presence of complex and high velocity data. The potential for reducing the incidence

of diagnostic errors in radiology must be addressed at all stages of radiology workflow,

including the final report, to have maximum impact on patient care (Bruno et al., 2015).

AI has the potential to assist and augment the radiologist during interpretation through its

capacity to access, analyze, and correlate data acquired from numerous sources (Dreyer

& Geis, 2017). This includes prior imaging studies, electronic health records, laboratory

studies, genetic profiles, published literature, and disease registries. AI can be used to

pre-analyze data, flag abnormal presentations, and subsequently prioritize the order of

what needs to be interpreted by the radiologist (Kahn, 2017). A single health care

provider such as a radiologist or pathologist cannot keep track of all relevant information

on a particular patient or on a disease process. AI solutions can help overcome human

limitations such as lack of experience or unawareness of possibilities.

Radiology workstations of the future will be required to meet new standards based

upon performance, reliability, and scalability. Effective performance will require easy

access to an AI menu of options, technological interoperability, and seamless networking

with external resources (Kahn, 2017). Adaptive workstations will be used by radiologists

to leverage the experience of peers from remote locations and to democratize expert

decision support with the help of AI. It will also be used by the radiologist to acquire

knowledge-based assistance from databases and published resources almost

simultaneously. In summary, for radiology workflow to be successful it must support the

role of the radiologist and lead to better patient care.

Natural Language Processing

The science of linguistics encompasses the form, meaning, and context of sounds

and symbols used to communicate (Zipf, 2012). Natural language processing (NLP)

represents a solution for linguistic analysis. It can be used to assign meaning or value to

unstructured (textual) data allowing it to be analyzed with computational methods (Pons,

Braun, Hunick, & Kors, 2016). The majority of digitized health care data within a

patient’s medical records and within the realm of population databases is unstructured.

The presence of pathology often reported semantically in radiology reports (Acharya et

al., 2017). NLP systems can access prior pathology and radiology reports to reveal prior

evidence and characteristics of pathology. NLP can therefore be used to leverage

information within electronic medical record systems to generate an active problem list.

It can also be used to provide access to prior imaging study findings for comparative

analysis and to provide decision support during radiology workflow (Massat, 2018).

Biomedical text mining with NLP leads to knowledge discovery that can aid imaging

interpretation.

NLP uses different methods to analyze unstructured data. The fundamental

mechanisms include pattern matching, parsing, and statistical approaches (Cai et al.,

2016). Hassanpour and Langlotz (2016) demonstrated that ML combined with NLP could

be effectively used to analyze textual data acquired from medical records including prior

radiology reports to assist in the differential diagnostic process. NLP represents an

important data management tool within the interpretive stage of radiology workflow. AI-

supported NLP can only reveal what was previously described, not the current state of

pathology evaluated through imaging. Characterization of an individual’s current state of

pathology requires in vivo imaging and quantitative measures using various techniques

such as radiomics. The combined use of NLP and radiomic methods will improve the

radiologists’ access to relevant information.

Radiomics and Radiogenomics

Radiomics refers to the science associated with high-throughput extraction and

quantitative analysis of non-visible data acquired from images to characterize pathology

(Lambin et al., 2017; Lee et al., 2017; Limkin et al., 2017; Griethuysen et al., 2017; Yip

& Aerts, 2017). Radiomics, first reported in 2012, has evolved quickly during the last

couple of years (Verma et al., 2017). Automated radiomic functions include pathology

detection, segmentation, feature extraction, and feature analysis (Lambin et al., 2017;

Limkin et al., 2017; Zaharchuk et al., 2018). Radiomic methods can be used to expand

available data within three-dimensional space on imaging studies to characterize

pathology and subsequently refine the diagnostic process (Gillies et al., 2015; D. Kumar

et al., 2012; Lee et al., 2017; Lambin et al., 2017; Limkin et al., 2017; Peeken et al.,

2018). Expanded dimensionality amplifies spatial heterogeneity (Cook et al., 2014). One

of the goals in radiomics is to identify disease features and biomarkers that have greater

causal rather than correlative relationships (Sanduleanu et al., 2018). Radiomic methods

have the potential to expand sub classifications of disease and support more personalized

care.

Radiomic measures can be used to quantify characteristics of pathology that are

not visible to the radiologist. This form of analysis addresses features of pathology such

as shape, volume, edge characteristics, texture, and other statistical measures (Aerts,

2016; Court et al., 2016; Gillies et al., 2016; V. Kumar et al., 2012; Lambin et al., 2012).

Radiomics is a discovery process rather than a validation process performed using one of

two approaches. The first approach focuses on mining images for predetermined patterns

of disease. The second approach uses deep learning methods to discover and learn disease

features not currently known. Radiomic methods like other analytic approaches tested for

reliability and validity prior to clinical use. Testing should include evaluation of

sensitivity and specificity along with the reliability of predicting positive and negative

values. The testing must be disease-specific and take into account different cohorts. The

primary goals of radiomic development are to help detect, characterize, and monitor

pathology. This includes classifying and staging disease.

Radiomic methods have been successfully used to detect and characterize some

diseases using different imaging methods such as CT, MRI, and PET (Acharya et al.,

2018; Parekh & Jacobs, 2016). Radiomics has been successfully used in different

research settings to evaluate breast cancer (Li et al., 2016), brain tumors (Li et al., 2017),

lung disease (Bak et al., 2018; Vallieres et al., 2015), liver disease (Naganawa et al.,

2018), brain metastasis (Ortiz-Ramon et al., 2017), and prostate cancer (Tanadin-Lang,

2018). Most of the important contributions to radiomics have come from the field of

oncology (Gillies et al., 2016). Radiomic methods show promise in the field of breast

imaging for the differentiation of benign versus malignant tumors (Hui et al., 2016).

Specific radiomic features such as enhancing tumor volume and texture features are

emerging as discriminatory factors in the differential diagnostic workup of breast cancer

(Drukker et al., 2018; Hui et al., 2016). Ongoing research is required to identify the

potential applications and benefits of radiomics in different fields. Future research will

reveal whether some of the methods used in other fields may be adapted and used in

spine care.

Radiomic methods are not limited to the initial diagnostic process. They can also

be used to help monitor disease progression and the response of pathology to treatment

(Vargas et al., 2017; Zhou et al., 2017). Radiomic methods can be used for the

surveillance of early-stage pathology prior to intervention. In addition, radiomic

signatures can ”…be used as precision biomarkers for the prognosis of individual

patients” (Castiglioni & Gilardi, 2018, p.412). The clinical role of radiomics is based on

the premise that imaging of disease requires a large volume of data that reflects multi-

scale pathological mechanisms, not detectable through routine visual assessment of

images.

The specialized field of radiogenomics sometimes referred to as imaging

genomics refers to the science associated with the correlation of anatomic characteristics

(phenotype) of pathology with genetic (genotype) data (Pinker et al., 2017).

Radiogenomic methods have been successfully used in the evaluation of lung cancer,

glioblastoma multiforme, kidney cancer, prostate cancer, and liver cancer (Incoronato et

al., 2018). Radiomic measures have proven useful for revealing phenotypic

manifestations of genetic expression (Giger, 2018; Panth et al., 2015). Phenotypic

characterization is important for there are many non-genetic determinants in pathology,

especially involving age-related diseases (Oakden-Rayner et al., 2017). Radiogenomics

has the potential to support precise classification of disease and subsequently the

personalized delivery of care (Bai et al., 2016). This conclusion is based on the premise

that alterations of genetic expression influence pathology represented by phenotypes

revealed through diagnostic imaging methods.

Volumetrics: New Realities and 3D Perspectives

Radiologists have a unique opportunity to combine the use of AI and volumetric

datasets to display pathology in three dimensions (3D) to help inform the diagnostic

process and therapeutic planning (Farahani et al., 2017). A few years ago, Denzel et al.

(2014) acknowledged the importance of in vivo interrogation of pathology within the

framework of volumetric imaging and display. A study limited to planar or 2-D imaging

can limit the diagnostic process.

Advanced imaging methods such as MRI and CT are capable of acquiring high-

resolution volumetric data sets that can be formatted to create 3D perspectives (Douglas

et al., 2016). The use of augmented reality and 3D viewing can improve the conspicuity

of pathology features (Douglas et al., 2016; Hamacher et al., 2016). Segmented pathology

can be manipulated and interrogated using virtual cut plane technology.

Multidimensional imaging data sets can be used to create virtual reality (VR) and

augmented reality (AR) based representations of pathology. Virtual reality offers an

immersive experience that can be used for planning and training. In contrast, augmented

reality offers digital images or prompts in the physical world. AR can be used to project

nonvisible perspectives of pathology during an exploratory biopsy or over a surgical

field, thus, limiting attention shifts between available imaging and the patient. The

augmented virtual or interactive display of pathology may soon be used to guide the

digital (virtual) biopsy and acquisition of radiomic measures.

Disease Modeling and Computational Diagnosis

Mathematical formulas can be used to help study normal biological and disease

processes (Mapoka et al., 2013). In silico modeling sometimes referred to as ex vivo

modeling refers to the use of quantitative imaging data to create comprehensive

mathematical models which can be used within the context of a computer system to study

the characteristics and behavior of pathology (Jeanquartier, Jean-Quartier, Cemernek, &

Holzinger, 2016; Louis et al., 2016; Mapoka et al., 2013). Mathematical models support

the use of simulated variables to predict the behavior of disease under different

circumstances (Louis et al., 2016). Radiomic features acquired in vivo can be used to

help develop in silico models (Echegaray et al., 2016). The use of mathematical disease

models built from data acquired over time through monitoring of a disease process could

help support predictive modeling.

The benefits associated with in silico modeling include the ability to study

disease, zoom in on pathology subsystems, and manipulate time scales, all of which

reduce the need for laborious and costly biological experimentation (Jeanquartier, et al.,

2016). Computational disease models can be used to test assumptions and hypotheses.

They can also be used to help predict outcomes and to reduce the frequency and

magnitude of uncertainties surrounding disease behaviors (Mapoka et al., 2013). Access

to mathematical disease models from the radiology workstation can augment the

differential diagnostic process. In the future, mathematical models will be used to help

identify ground truth and to train AI systems.

Computational diagnostics is the science associated with the combined use of

algorithms, mathematical formulas, and computer systems to detect and study disease.

Computational pathology refers to the use of a digital disease model to evaluate

assumptions and to study disease behavior under simulated circumstances. The

convergence of interoperable technologies offers an unprecedented opportunity to

correlate pathomic, radiomic, and genomic features to create computational disease

models. In the near future, computational analysis may become as important to radiology

as the microscope is to pathology (Louis et al., 2016). Access to AI supported

computational diagnostic tools during the interpretive stage of radiology workflow will

enable radiologists to address the complexities of pathology.

Workflow Optimization and Data Integration

The differential diagnostic process in radiology typically requires correlation of

imaging findings with data from other sources such as prior imaging studies, medical

records, peer-reviewed literature, and disease registries. Radiology workstations will

eventually consist of embedded AI solutions and networked functions to perform

interpretive functions, to obtain second opinions, and to support multidimensional

viewing (Morgan, Branstetter, Mates, & Chang, 2006). Radiomic workflow includes data

acquisition, pathology segmentation, feature extraction, feature analysis, and correlation

with computational disease models. The radiologist will serve as the expert on the in vivo

diagnostic process and as a gatekeeper to the flow of data and access to knowledge from

diverse data sources.

The radiology workstation of the near future will likely represent the principle

hub of converging data from the fields of radiology, pathology, and genetics (Gillies et

al., 2016; Jha & Topol, 2016). Computational methods will be required to analyze data

from these diverse sources. Unlike the radiologist, AI solutions do not represent a single

layer of interpretation. It represents multilayered approaches, consisting of logical

analysis and problem solving, not subject to the variability and inconsistencies of a

biological system.

Data acquired from nonimaging sources need to be annotated or tagged in a

manner that prepares it for sorting, aggregation, and analysis. Complex data often has to

be conditioned and prepared for exposure to algorithms and computational analysis. The

curating process requires well-defined steps to identify inconsistencies, to standardize

symbolic representations, and to transform data to a uniform interoperable and

interpretable language. This process can be augmented through the identification of

common data elements (CDEs). CDEs help integrate disparate clinical, phenotypic, and

genotypic data. The consistent use of CDEs supports the development and use of disease

registries (Rubin & Kahn, 2017). Widespread application of CDEs will support the

development of integrated knowledge databases accessible by AI for use during the

interpretive stage of imaging. Curated data sets are required to train AI.

Access to Disease Databases and Registries

The increasing ability to extract pathology features during imaging procedures

will result in new forms of data storage and disease registries (Rastegar-Mojarad et al.,

2017). Digital tissue samples can be stored in a manner that will maintain accurate spatial

and relational data. Unlike in vitro tissue samples, in vivo digital representation of

pathology will not degrade. Digital representation of in vivo characteristics can be

acquired and stored in a manner that maintains three-dimensional relationships. The

radiologist will have increasing access to digital pathology registries and mathematical

disease model libraries from their workstation to augment the differential diagnostic

process.

Structured Reporting

The radiology report reflects what took place during the interpretive stage of

radiology workflow. The decisions made influence the content and structure of the final

report. Structured reporting refers to the use of a standardized format, quantitative

measures, and consistent descriptions. Structured radiology reports can be

operationalized with quantitative measures derived from AI applications (J. Y. Chen et

al., 2017; Dreyer & Geis, 2017). Growing use of widely accepted terminology and

standardized disease nomenclature will support more consistent diagnostic reporting and

reliable NLP applications. Using keywords and phrases on reports can serve as common

data elements used to trigger prospective or retrospective interpretive or comparative

processes. For example, describing focal pathology in a prior radiology report may direct

auto detection and quantitative radiomic measures within the same volume of interest in a

subsequent study. Structured reporting provides more concise decision support at the

point of care (Alkasab et al., 2017). According to Schwartz et al. (2011), “Referring

clinicians and radiologists found that highly structured reports had better content and

greater clarity than conventional reports” (p. 174).

In the near future, AI will offer a full spectrum of solutions that will begin with

pathology detection and end with automated reporting within a structured format

(Zaharchuk et al., 2018). The combined application of ML techniques and NLP to extract

patterns from current imaging studies and prior radiology reports will improve imaging

interpretive accuracy (J. Y. Chen et al., 2017). Text clues within a prior radiology report

can be used to trigger access to current imaging and prior non-imaging data to support the

interpretive process.

Digital Exploration: The Virtual Biopsy

Growing use of AI combined with radiomic methods will expose new features of

disease and expand the spectrum of pathology. This process will lead to a longer list of

disease subtypes and differential diagnostic possibilities during the interpretative stage of

imaging, thereby, increasing the complexity of decision making. Advances in technology

will support the exploration and interrogation of in vivo pathology with visual,

augmented, and subvisual techniques. Augmented visual approaches include the use of

interactive 3-D displays of pathology.

Subvisual Tissue Interrogation

Radiomic methods can be used to detect and characterize features of pathology

from imaging studies in a manner undetectable by traditional visual interpretation (Aerts,

2017; Gillies et al., 2016). Advanced algorithms can be used to correlate subvisual

pathology features with other sources of data such as histological features from pathology

and genetic profiles (Bucking et al., 2017). Echegaray et al. (2016) introduced the

concept of the “digital biopsy” which refers to the targeted non-invasive acquisition of

pathology features in vivo” (p. 283). In another field, Mancini et al. (2018) introduced the

concept of the “digital liver biopsy” referring to the use of multimodality and

multiparametric imaging of the liver (p. 3). The use of a digital or virtual biopsy approach

has the potential to interrogate and map the entire landscape or volume of a pathological

state, whereas the traditional needle biopsy offers limited characterization of pathology

(Echegaray et al., 2016; Lambin et al., 2012; Thrall et al., 2016). It is important to

evaluate and characterize a whole region of pathology to reduce or eliminate sampling

bias. Under sampling may poorly represent the entire state or spectrum of pathology

present and subsequently misdirect treatment. Traditional and virtual biopsy results can

be combined to better characterize pathology. The virtual biopsy can be also used to

augment or direct the traditional needle biopsy.

Radiomics is not limited to the detection and extraction of pathology features. It

can be used to create or reveal new quantitative descriptors of pathology. Subsequently,

radiomic methods will continue to contribute to the development of new molecular

biomarkers and imaging signatures of pathology (J. Wu et al., 2018). For this reason,

radiomic methods are being considered for the expansion of the traditional tumor-node-

metastasis (TNM) staging process in oncology (Lai-Kwon, Siva, & Lewin, 2018). Digital

exploration of pathology can be enhanced through partitioning of an image and by

digitally extracting tissues or structures from the field of view, which can enhance

regions of interest. Special digital filters can also be used to increase the dimensionality

of a volume of interest (VOI), thereby improving tissue feature analysis and the ability to

sub-classify disease (Parekh & Jacobs, 2016). Edge detection algorithms is used to help

identify the boundaries between healthy and diseased tissues. Contour analysis is used to

segment and highlight the borders of pathology. Knowledge of the true border of

pathology supports more accurate volumetric measures and feature mapping of

pathology. Radiomics is capable of providing an automated solution for the assessment of

disease characteristics and evolution.

Radiomic methods do not always serve as the final or conclusive method of

pathology assessment. They can be used to help identify the signature for an aggressive

region of pathology that can serve as a high-risk target for a traditional needle biopsy

within a well-defined region of pathology (Sala et al., 2017). Radiomic feature analysis

using multiparametric or hybrid imaging technologies such as PET/CT and PET/MRI can

be used to interrogate large volumes of tissue in vivo with subcentimeter resolution,

which enhances the accuracy of the virtual biopsy (Kressel, 2017). Integrated data from

these approaches can also be used to better guide traditional biopsy methods.

Pathology Feature Extraction and Analysis

Radiomic methods are effective for revealing geometric, statistical, and textual

features of pathology from imaging data. Over 440 radiomic features of pathology have

been acknowledged in the literature (Wu et al., 2018). Common categorical features of

pathology include texture, edge-attributes, volume measures, molecular relationships,

contrast kinetics, uptake values, and subvisual pathology boundaries (Bai et al., 2016).

Individual voxels and three-dimensional matrices comprised of an array of voxels

provide the volumes of interest (VOI) for radiomic assessment. Texture analysis refers to

the quantitative evaluation of grayscale intensities and their relationships within and

between well-defined three-dimensional space referred to as voxels (Davnall et al., 2012;

Lubner et al., 2017). Texture may correspond to various pathological processes.

Specialized digital filters can be used to help reveal unique features of pathology such as

entropy and uniformity (Nandu, Wen, & Huang, 2018). The greater the number of

clinically relevant pathology features, the greater the potential for differentiating and

classifying pathology.

Diagnostic images are comprised of first, second, and third order data which when

exposed to advanced computational methods can detect patterns, relationships, and trends

that support health care decisions (Gillies et al, 2016; Huang et al., 2017). First-order

statistics include mean gray-level intensity, standard deviations, entropy, skewness,

kurtosis and uniformity. Second-order statistics include local homogeneity, dissimilarity,

correlation, and angular second moment energy (Gillies et al., 2016). Higher order

statistics include coarseness, contrast, and complexity (Davnall et al., 2012). Data can be

acquired and analyzed during the interpretive stage of radiology workflow with

handcrafted and/or deep learning radiomic methods.

Multidimensional 3D Exploration

A focal region of pathology such as a tumor often has a high degree of spatial and

temporal heterogeneity, which limits the usefulness of conventional structural imaging

and traditional biopsy results. A digital (virtual) biopsy can be performed using an in vivo

voxel by voxel (voxel-wise) interrogation process across the entire volume of pathology

in multiple dimensions. Voxel-wise classifiers can be used to help reveal varying stages

and subtypes of pathology within a single lesion (Bucking et al., 2017; Ng et al., 2013).

Radiomic methods have been successfully used to help differentiate benign and

malignant characteristics of tumors and to provide prognostic insights (Sala et al., 2017).

Further research will expand knowledge of the molecular attributes and radiomic

signatures of aggressive pathology within different biological systems and tissues.

Comprehensive in vivo assessment of pathology requires analysis across the

three-dimensional volume of pathology rather than from a two-dimensional plane

(Echegaray et al., 2016; Nandu, Wen, & Huang, 2018). As data acquisition and

computational methods become faster, the trend will be to auto interrogate larger

volumes of pathology and auto map pathology features using in vivo data (Zhao et al.,

2016). Virtual exploration supports interrogation of pathology within an augmented or

immersive 3-D environment.

Pathology does not exist in isolation. It functions and interacts within a biological

ecosystem that includes the surrounding microenvironment. Within this ecosystem, there

is both phenotypic and genotypic plasticity that contributes to the evolution of pathology.

The use of multiparametric imaging methods and radiomic measures can help expand and

reveal feature data sets from the whole region of pathology including the surrounding

microenvironment. This represents a significant advantage of multidimensional

interrogation of pathology. It also offers an advantage over lab (blood) studies that

represent circulating biomarkers, which do not localize pathology within an organ or

tissue. In the near future, whole pathology slide mounts will be matched with image slice

acquisitions and voxel by voxel radiomic measures. This approach will be used to help

build computational disease models and disease detection algorithms.

Disease Screening: Discovery Radiomics

Early detection and characterization of pathology influences patient care and

treatment outcome. Handcrafted algorithms and rules used with AI are often limited in

their capacity to reveal subtle or unknown characteristics of healthy and disease states. D.

Kumar et al. (2017) introduced the concept of “discovery radiomics” which refers to the

use of high-throughput analysis of imaging data to detect early-stage pathology and to

help predict the outcome of an asymptomatic disease process. A specialized approach to

data analysis referred to as “evolutionary deep radiomic sequencing” offers a rapidly

evolving method for detecting pathology, which is not dependent on a priori knowledge

of disease criterion (Shafiee et al. 2017). In summary, various forms of radiomic

applications can be used to provide a low-cost, fast, and potentially reliable method of

screening for pathology.

Reimagining the Differential Diagnostic Process

The differential diagnostic process is comprised of a series of interrelated steps

applied to a patient’s presentation, which uses probability-based logic or reasoning to

differentiate a disease or disorder from others that may have a similar presentation. The

differential diagnostic process is often dependent upon different sources of information

such as the history, physical examination, laboratory evaluation, imaging, and other

specialized forms of testing. In general, the more complicated a patients’ presentation, the

greater the list of possible causes and contributing factors. Advances in diagnostic

imaging have contributed to a growing appreciation for the complexity and heterogeneity

of disease at anatomic, cellular, molecular, and genetic levels (Aerts, et al., 2013; Davnall

et al., 2012; Lubner et al., 2017; Sala et al., 2017; Yip & Aerts, 2017). Radiomic

measures can be used to “ bridge evidence across different biological scales” in a manner

which can inform the differential diagnostic process (Hsu, Markey, & Wang, p. 1010). It

is important to discover new dimensions and features of pathology that have meaningful

impact on patient care.

Heightened awareness of the complexity of disease expands the list of differential

diagnostic considerations. Advances in diagnostic imaging will continue to reveal unique

heterogeneic features that can be used to classify and subtype pathology (McCue &

McCue, 2017). The heterogeneity of pathology is associated with evolutionary changes at

the cellular level associated with genotypic and phenotypic plasticity. Heightened

awareness of this process improves diagnostic and prognostic accuracy. For example,

high levels of tumor heterogeneity have been associated with greater variability of

treatment outcome and generally a poorer prognosis (Lundstrom, Gilmore & Ros, 2017;

Sala et al., 2017). Successful delivery of more precise and personalized health care

requires knowledge of biological differences and pathological variability along with

relevant decision support at the radiology workstation.

The current diagnostic process is very nuanced, and influenced by a provider’s

familiarity with disease and related testing. In addition to assisting with disease detection

and characterization, AI can be used to identify appropriate diagnostic tests and related

protocols to help achieve a precise diagnosis (Baldwin, Guo, & Syeda-Mahmoood,

2017). There are differential diagnostic possibilities for every patient presentation

(Hussain & Oestreicher, 2017). The individual health care provider typically relies on an

intuitive diagnostic approach limited to their familiarity of two to six diseases during the

initial differential diagnostic process (Phua & Tan, 2013). This level of awareness is

often insufficient for unusual or complex conditions. Limited awareness of differential

diagnostic possibilities leads to errors and missed treatment opportunities.

Traditionally, radiology has relied upon the visual perception of the radiologist,

which limits their capacity to consider microscopic and molecular level differential

diagnostic considerations (Pinto & Brunese, 2010; Yip & Aerts, 2016). Radiomic

methods can be used to overcome visual limitations and expose measurable

characteristics of subvisual pathologic states, thereby, adding value to the differential

diagnostic process. Improvement of the differential diagnostic process is required in all

areas of diagnostic imaging including spine care.

The application of new quantitative imaging (QI) measures and standards will

refine the diagnostic process and lead to expanded criterion and classifications of disease

(Farooki et al., 2016; Gillies et al., 2016; Kharat & Singhal, 2017; V. Kumar et al., 2012;

Parekh & Jacobs, 2016). The expansion of objective measures of pathology will support

AI solutions. Disease states and their evolution vary between individuals because of

molecular, genetic (genotypic), structural (phenotypic), and biological diversity. For this

reason, a one-size-fits-all approach to diagnosis or treatment does not work for everyone.

Radiomic analysis, a specialized application of QI, can be used to characterize biological

variability and pathological heterogeneity at subvisual levels independent of the

radiologist’s interpretation (Larue et al., 2016; Yip & Aerts, 2017). Radiomic data will

assist the radiologist in the differential diagnostic process. The adoption of QI will also

support further development of AI by providing objective labels used to annotate training

and validation data.

The term probability in the diagnostic process refers to measures of the likelihood

of a disease or pathologic process being present. AI can expose radiologists to differential

diagnostic considerations (diseases) for which they have limited knowledge and

experience. Probability is assigned to disease patterns and differential diagnostic

possibilities associated with imaging presentations derived from hundreds of thousands of

patients within a database or from case-based publications.

In summary, treatment success is dependent upon an accurate and efficient

differential diagnostic process. The current diagnostic process remains too imprecise and

inconsistent to adequately identify and subtype disease in many situations. This often

results in a “one-size-fits-all” treatment approach. AI solutions have the potential to use

probability calculations to help identify and prioritize diagnostic possibilities. AI will

continue to influence the steps, as well as the sequence of steps used during image

interpretation and the differential diagnostic process. The combined use of AI-supported

data management solutions such as natural language processing, radiomic methods,

machine learning, and advanced computational analysis at the radiology workstation

might support improved accuracy and precision of the differential diagnostic process.

The Radiologist: New Roles and Responsibilities

The principal role of the radiologist is to detect, characterize, and report on

disease processes in a manner that provides decision support at the point of care. For this

reason, there is a growing demand for radiologists to become more involved in

consultation and patient care (Ranschaert, 2016). Historically, radiologists are skilled in

the evaluation of pathology associated with structural changes and are less skilled in the

early detection of pathology based on subvisual criteria revealed by deep learning and

radiomic methods (Malone & Newton, 2018). This inadequacy reinforces the need for

computational assistance. AI has the potential to augment the role of the radiologist

during the interpretive stage of radiology workflow and with final reporting. More

specifically, AI may improve the performance of the radiologist by enabling earlier

disease detection, offering more precise disease characterization, providing probability

based differential diagnostic considerations, and reducing diagnostic error rates

(Zherhoni, 2017). Radiologists are rapidly becoming among the most important data

managers, knowledge brokers, and primary gatekeepers of big data and curators of

knowledge for treatment planning and disease surveillance in health care (Hillman &

Goldsmith, 2011; Jha & Topol, 2016). Using AI during the interpretive stage of radiology

workflow can help identify meaningful incorporated into structured reports.

Despite these important roles, the interpretation of diagnostic images is still

primarily limited to the visual detection and characterization of pathology (Pinto &

Brunese, 2010). This approach is no longer adequate. In some areas of radiology AI has

proven it has the potential to augment the role of the radiologist in the detection and

interpretation of subvisual data (Farooki et al., 2016; Larue et al., 2016). AI has the

potential to empower the radiologist to work better, faster, and smarter.

AI can augment the role of the radiologist and empower their role as a clinical

consultant in numerous ways. For example, AI can help access, analyze, and correlate

non-imaging data with imaging findings during the interpretive stage of radiology

workflow. AI can also be used to flag subtle or visually hidden pathology and address

mundane and redundant work, thus, freeing radiologists up to interpret pathology and

better communicate with referring physicians and other members of the health care team

(Recht & Bryan, 2017). AI may overcome human limitations such as fatigue, lack of

experience, unawareness of possibilities, and bias. To maintain relevance the radiologist

must be willing to embrace AI, adapt to its use, and contribute to its development.

Additional research is required to address how AI applications augment the role of the

radiologist and improve the delivery of more precise and personalized care (Sutton et al.,

2017).

Artificial Intelligence in Radiology: Current Applications

The individual radiologist is often overwhelmed with the growing burden of high

volume complex data. This dilemma has led to the pursuit of different forms of decision

support. This includes AI solutions. Most of the research surrounding the use of AI in

radiology has been limited to highly specialized fields. There has been little research

surrounding its role in spine imaging. Successful non-spinal applications will pave the

way for use in spine care.

Non-Spinal Imaging

Because of the complexity of available data and the criticality of decisions, most

research on how AI is used in radiology is limited to cardiology, neurology, and

oncology. AI solutions are used in other specialties of radiology albeit to a limited

degree. Deep learning methods have been successfully used to auto detect pulmonary

tuberculosis on chest radiographs (Lakhani & Sundaram, 2017) and to auto detect disease

states in neuroimaging such as intracerebral hemorrhage, stroke, and mass effects (Maier

et al., 2015; Prevedello et al., 2017; Scherer, 2016). Computer-aided diagnostic systems

help detect and characterize some neurodegenerative disorders (Cascianelli et al., 2016).

AI has evolved to a level where it has outperformed the human expert in radiology in

some settings (Augimeri et al., 2016; Boone et al., 2015; Mohebian et al., 2017).

Research addressing the use of AI has demonstrated its potential to augment the role of

the radiologist.

The utility of AI is not limited to radiology. For example, deep learning

algorithms have demonstrated greater accuracy than a panel of pathologists in the

detection of lymph node metastasis in women with breast cancer (Bejnordi, Veta, & van

Diest, 2017). Extensive research is also underway to develop methods for automatically

extracting relevant information from unstructured reports in breast imaging (Gupta,

Banerjee, & Rubin, 2018). In another study, convolutional neural networks outperformed

cardiologist’s interpretation of echocardiographic images with 98% accuracy (Mandani et

al., 2017). Successful use of AI solutions in one field of radiology may be adapted to

meet the needs in another field such as spine imaging. The adoption of AI solutions

requires adequate research to identify meaningful applications, as well as to confirm

clinical utility and validity.

Spine Imaging

Diagnostic imaging is often a fundamental and influential component of spine

care. Imaging findings influence decision making at all levels of care. Subsequently,

interpretive imaging errors and missed opportunities have a profound impact on treatment

planning and treatment outcomes. An extensive literature search revealed limited

research and real-world applications of AI in spine imaging and spine care. It is common

for early applications of AI to be applied to basic steps such as anatomic localization and

labeling prior to implementation of more detailed applications. For example, machine-

learning algorithms with different imaging modalities can auto-identify vertebral levels

(Daenzer et al., 2014; Hetherington et al., 2017). Another research study demonstrated

that AI could be used to segment and label vertebral bodies (El-Helo et al., 2013). Auto

segmentation and labeling of anatomic regions has to be perfected before tissue features

can be auto extracted and characterized.

AI use in spine care has not been entirely limited to anatomic localization.

Machine learning methods have also been used to determine bone density, as well as to

detect and categorize vertebral compression fractures on computerized tomography

(Burns, Yao, & Summers, 2017; Doi, 2007; Hetherington et al., 2017). In another study

El-Helo et al. (2013) demonstrated that AI performed with greater than 90% accuracy in

the detection of vertebral compression deformities. This condition is relatively common

and places a significant financial burden on the health care system. Vertebral

compression deformities often missed on non-spinal imaging studies could be detected

with AI supported methods. For example, N. Kim et al. (2004) reported that

approximately 50% of vertebral compression fractures that presented on routine lateral

chest radiographic studies were either missed or underreported.

A few research studies have exposed the potential utility of molecular imaging in

spine care. For example, in vivo quantitative voxel-based mapping is used to evaluate the

microstructural and molecular attributes of degenerative intervertebral discs and of the

spinal cord in cervical spondylotic myelopathy (Grabhar et al. 2015; Grunert et al., 2014).

AI could be used to help identify the presence of subtle spine pathology that may be

missed during non-spinal imaging studies, which include spine data in the background.

Most diseases progress through an asymptomatic (subclinical) period. For

example, early spinal cord compromise (myelopathy) secondary to degenerative stenosis

and compression often results in asymptomatic changes in regional biochemistry, blood

flow, and tissue architecture preceding the onset of clinical signs and symptoms (Durrant

& True, 2012). These changes are often not evident on routine imaging studies. The

visual presence of pathology within the spinal cord on advanced imaging studies is often

associated with end-stage pathology and permanent neurological deficits. Successful

detection of early stage myelopathy will require non-visual analysis of data acquired form

the spinal cord with molecular imaging and AI-supported radiomics. Researchers have

begun to address the possibilities. For example, a specialized form of MRI referred to as

diffusion tensor imaging (DTI) combined with machine learning classifiers has been

successfully used to detect early stage spinal cord compromise secondary to degenerative

narrowing of the central spinal canal in the neck, a condition referred to as cervical

spondylotic myelopathy (Wang et al., 2015; Wang, Hu, Shen, & Li, 2018). The concept

of discovery (screening) radiomics can applied to any bodily region including the spine

and spinal cord.

Computational AI approaches have been used to auto classify intervertebral discs

as either normal or degenerative based upon the analysis of tissue features such as signal

intensity, texture, and shape (Ghosh & Chaudhary, 2014; Oktay, Albayrak, & Akgul,

2014). In another study, AI assisted the automated detection and characterization of

lumbar neuroforaminal stenosis on MRI studies with greater than 90% accuracy (Han et

al., 2018). Further research is required to determine how AI-supported methods could

augment the role of the radiologist during the interpretive stage of spine imaging. Further

research on the use of AI in radiology will shape the future spine care.

Spine Pathology and AI: Meaningful Use Considerations

The capabilities of deep learning AI systems are highly dependent on exposure to

adequate levels of annotated training data and the establishment of ground truth. The

process is often expensive, tedious, and time-consuming. Due to this level of

commitment, it is imperative for stakeholders, including clinicians and radiologists, to

help identify meaningful use applications prior to investing in the process. Meaningful

use in this context refers to the application of an AI solution to address a condition or

disease which is prevalent, not adequately assessed with normal methods and which has a

significant impact on the individual and society. The institutional definition of

meaningful use applications in this context may include economic value assigned to new

solutions. An example of successful AI application with a favorable outcome is referred

to as a meaningful use case.

Some spine disorders, if left undetected and untreated, can lead to devastating

consequences that place an unnecessary burden on the individual, their family, and

society. The evaluation of AI use in spine imaging should start with the most devastating,

prevalent, and costly spine disorders. Examples include the 550,000 to 700,000 vertebral

compression fractures which occur annually in the United States secondary to

osteoporosis (Kondo, 2008) and the two-thirds of patients with cancer who will develop

bone metastasis, with the spine representing the most common location (Maccauro et al.,

2011). Intervertebral disc degeneration represents one of the most common causes of

back pain, which afflicts approximately 80% of the adult population throughout their

lifetime (Suthar et al., 2015). The annual prevalence of spinal cord compromise

(myelopathy) secondary to degenerative changes is estimated to be about 196,000 in

North America (Nouri et al., 2015). Each of the above conditions occurs in well-defined

anatomic regions of the spine, which can be interrogated in vivo at multidimensional

levels using emerging AI-supported methods

Some of the most easily segmented and labeled anatomic regions of the spine

house some of the most devastating types of pathology. These regions include the bone

marrow microenvironment within the vertebral body and the spinal cord

microenvironment within the spinal canal. In each case with the exception of trauma,

severe pathology begins with nonvisible, asymptomatic, tissue changes. Early detection

and intervention may lead to better patient outcome. AI-supported interrogation of

vulnerable anatomic regions could help detect and characterize aggressive pathology and

lead to early-personalized intervention.

The bone marrow microenvironment within the vertebral body is involved in

many different disease processes. For example, changes within the bone marrow often

precede the development of degenerative disc disease, as well as the development of

compression deformities and fractures. The bone marrow space is also a common

location for metastatic disease. Approximately two thirds of individuals with cancer will

develop bone metastasis with the spine representing the most common site (Maccauro et

al., 2011). Micro metastasis or early metastatic disease is often subclinical and difficult to

detect with traditional anatomic imaging methods. Approximately 36% of vertebral

lesions associated with spine metastasis are asymptomatic and discovered incidentally on

spine imaging for other disorders (Maccuaro et al., 2011). In some cases evidence of

metastasis to vertebral bone marrow may represent the first indicator that cancer is

present somewhere else.

The vertebral bodies and bone marrow are well visualized on routine spine MRI

and CT studies; therefore, rendering it possible for AI supported screening methods to be

applied to detect subtle or early stage pathology. Prior to implementing AI supported

solutions such as radiomics ground truth must be established for normal and abnormal

states. Discovery radiomics and in vivo interrogation of vertebral body

microenvironments may help reveal early stage osteoporosis, micro metastatic disease,

and subchondral degenerative changes that often precede the development of

intervertebral disc disease. Early detection may lead to early intervention.

Some advanced imaging methods and protocols have demonstrated the ability to

reveal micro pathology within the bone marrow of the spine (Long, Yablon, & Eisenberg,

2010; Park et al., 2015). AI-supported radiomic methods have demonstrated the ability to

detect nonvisible evidence of pathology in other fields. Its potential role in spine care

must be investigated further. In the future, radiomic methods will likely detect and

characterize vertebral bone marrow pathology such as myeloproliferative disorders,

subchondral degeneration, osteonecrosis, infection, and tumor (Long et al., 2010). It is

imperative to diagnose pathology within the vertebral bone marrow environment at the

earliest stage possible.

The first step in pathology detection using AI methods is auto assessment and

labeling of anatomic structures. The second step is to segment a region of interest. The

third step is to apply specialized AI applications such as radiomic methods to detect and

characterize pathology. AI has already been successfully used to perform automated

vertebral boundary detection and surface texture analysis using advanced context-

encoding features (Mirzaalian et al., 2013). This provides the digital framing and

segmentation required to isolate bone marrow. With the exception of one study, I was

unable to identify any significant research surrounding the use of radiomic methods to

evaluate tissues of the spine or spine pathology. In the referenced study, quantitative

voxel-based feature detection and analysis was performed using computed tomography

studies of the spine to assess true marrow space, fat composition, and mineral-based

marrow density (Pena et al., 2016). Pena et al. (2016) found that radiomic methods offer

advanced tissue differentiation and characterization that could be applied to the spine.

The same spine disease or disorder within different individuals may vary in

presentation, severity, and progression due to molecular, genetic, and structural diversity.

For this reason, a one-size-fits-all approach to spine imaging diagnosis or to treatment

may not work. AI-supported methods such a radiomics can help classify and stratify

disease and therefore help identify the best-personalized treatment plan. This research

study may represent one of the first to address the potential impact of AI in spine imaging

along with the potential role of radiomics and the virtual biopsy for the in vivo

interrogation of spine pathology.

Combinatorial Evolution of Artificial Intelligence

The pace of technological development and the rapid evolution of AI will

continue to increase. As technology advances the duration between innovations and new

applications often becomes shorter. In radiology, the phenomenon is amplified by co-

evolution of integrated technologies such as computer processing, imaging modalities,

database networking, and AI workflow solutions. As AI provides better analytic insights

and decision support there will be a greater push for more advanced imaging technology,

further complicating the decision-making process. Once human limitations are overcome,

there will a rising demand for the acquisition and analysis of more complex data to

support more precise and personalized care. Limited decision support for complex data

was previously a barrier to technology development and technology adoption in

radiology.

Perpetual revisions of disease criterion and classifications combined with ongoing

disease biomarker discovery will result in recurrent cycles of disruption and adaptation in

radiology and related clinical workflow. The three principal influences of an innovation

are differentiation, precision, and speed. Each of these factors will be pushed to the limits

in radiology by AI. There is a recursive relationship between technology, data, and

human contributions, all which collectively influence diagnostic decisions and the

delivery of care. The co-evolution of technologies at the radiology workstation

surrounding the management of high-throughput data will lead to unprecedented methods

of pathology assessment. AI and related technology support will lead to expanded

knowledge of disease pathophysiology and new standards of evidence-based care.

Exploratory research methods will help reveal the potential use of AI in radiology and

heighten awareness of and readiness for what is to come.

Imaging With AI: The Key to Precision and Collaborative Spine Care

Using big data and AI will transform the entire field of health care (Obermeyer &

Emanuel, 2016), including spine care. The use of AI solutions in radiology will forever

change how decisions are made in all health care specialties (Brink et al., 2017; Jha &

Topol, 2016; Jiang et al., 2017; Ranschaert, 2016). The unprecedented paradigm shift

associated with molecular level diagnostics and AI decision support will include spine

care. Personalized spine care requires identification of biological differences and

pathological variability, topics AI can help address.

Various forms of AI have the ability to improve the precision of the diagnostic

process in radiology by revealing unique attributes of disease (Syed-Mahmood, 2018;

Wang et al., 2015; Wang, et al., 2018; Zherhoni, 2017). For this reason, successful

applications of AI in radiology can have a favorable impact on the delivery of predictive,

pre-emptive and personalized health care (Augimeri et al., 2016; Brink et al., 2017; Jha &

Topol, 2016; Jiang et al., 2017; Ranschaert, 2016). The spine is intricate and complex.

The individual spine has many unique structural and biomechanical attributes that

influence the impact of disease. Greater knowledge of the unique attributes of an

individual’s spine pathology will inform all members of the spine care team. It will

support more precise and personalized spine care. Fundamental shared knowledge will

also facilitate more effective collaborative multidisciplinary care.

Growing use of AI in spine care will result in greater appreciation for the

spectrum of pathology, and for the need for a more timely and precise diagnosis. Future

AI applications in spine care will also expose common ground, redefine expert

boundaries, dampen existing professional turf wars, and lead to a few new ones.

Successful application of AI during interpretive stage of radiology workflow in any

specialty field of health care has the potential to transform the delivery of care while

supporting a more precise and personalized diagnostic process (Acharya et al., 2018;

Ghasemi et al., 2016; Hillman & Goldsmith, 2011; Jha & Topol, 2016). In addition to

improving the accuracy of the diagnostic process, AI may serve as a catalyst to new

levels of multidisciplinary collaboration in spine care. Some of the current applications of

AI in oncology, neuroradiology, and breast imaging will likely be adapted for use in

spine imaging.

A better understanding of the molecular basis of spine pathology using deep

learning and radiomic methods will lead to earlier detection and more precise subtyping

of pathology. This process will have an impact on the standard of care and will

perpetually reset clinical expectations. With heightened awareness of available decision

support radiologists, spine care providers, and the public will become less tolerant of

complacency and errors made during the interpretation of spine imaging. New

expectations and standards in spine imaging interpretation and reporting will have a

significant impact on evidence-based spine care.

The potential benefits of multidisciplinary care are well established. The success

of intervention is influenced by greater awareness of the heterogeneity of pathology,

acceptance of new disease criterion and classifications, and evolving standards in

decision support (Aerts, 2016; Collins & Varmus, 2015; Hood & Aufrey, 2013).

Collaborative spine care will become progressively more dependent on the central role of

the radiologist as a gatekeeper of big data and as a clinical consultant. The radiologist’s

role and impact is strengthened by improved diagnostic methods and better decision

support (Castenda, et al., 2015; Kressel, 2017). A more precise diagnostic process during

imaging workflow will help expose the fundamental basis of disease and subsequently

help bridge-the-gap between disciplines working at different points along the spectrum of

pathology (Brink et al., 2017; Kressel, 2017; Jiang et al., 2017). The collective use of

human and machine intelligence during the interpretive stage of spine imaging workflow

may help resolve multidisciplinary discordance by democratizing decision support and by

exposing standardized terminology and disease criterion.

Ethical Issues Surrounding the Use of AI in Radiology

Numerous ethical considerations surround the use of AI and big data in radiology

that could influence patient care (Mittelstadt & Floridi, 2016). Expanded use of AI will

subsequently require the creation of new ethical standards surrounding (Mesko, 2017).

Lee et al. (2017) raised concerns about potential challenges associated with deep learning

in radiology. The challenges included a growing dependency on large volumes of training

data and the black box nature of the technology. The later refers to the lack of

transparency of computational methods used to provide decision-support. The steps

associated with deep learning analysis of complex nested layers of computational

processes hidden. The lack of transparency limits validation and may lead to

unprecedented categories of liability.

Challenges associated with rapid adoption and evolution of AI in radiology

includes the possibility of over fitting results due to innovation bias, unnecessary hype,

and unrealistic expectations. Misdirected hype surrounding AI use in radiology could

have many direct and indirect adverse consequences on a health care system, such as

draining capital, unproductive reassignment of expertise, and the generation of false

expectations. An AI supported diagnostic process restricted to data mining can lead to

unproductive steps in workflow. Thus, Mayo et al. (2016) introduced the concept of

“farming of data,” in contrast to mining of data (p. 261). Data farming is characterized by

actively harvesting necessary data, locating missing data, picking the best data, and

weeding out unnecessary data. Ethical matters must be addressed to improve meaningful

use and the clinical utility of AI.

Widespread adoption of AI could result in deskilling of health care providers

including radiologists. In addition, the use of AI may result in a growing level of

technology codependency, thus, diminishing human influence in decision making.

Widespread AI adoption could also introduce automation and/or technology bias. Earlier

detection of pathology could result in unnecessary testing and treatment exposure. Access

to AI enhances authority and expertise and will afford the user with an advantage that can

have a significant impact on leadership roles, collaborative efforts, and the equality of

care. For the reasons stated, it is necessary to perpetually explore and address ethical

considerations that may arise because of the development and use of AI during the

interpretive stage of spine imaging workflow.

The Research Approaches

Researchers in the disciplines of computer science and radiology have approached

the potential role of AI in radiology from numerous perspectives. The approaches have

identified what may be missed with traditional anatomic imaging and qualitative

interpretation. Additional strengths are related to the investigation of narrow (disease

feature specific) applications of AI using methods such as radiomics and natural language

processing in isolated fields such as oncology. The weakness of this approach is the

limited knowledge acquired surrounding proposed AI, its interoperability with current

workflow, clinical utility, and ease-of-use.

There have been a limited number of qualitative exploratory studies on the

potential role of AI during the interpretive stage of radiology. The absence of qualitative

insights has resulted in numerous reductionist approaches to research on the topic with

limited capacity for generalization and practical clinical application. Many of the research

studies referenced in this work have not led to the development of an adequate concept

map or blueprint of the sequence of research required to address the potential role and

impact of AI on the interpretation of imaging studies and on the final reporting process.

Exploratory research will help identify needs and potential and reveal the sequence of

research studies and methodologies required to achieve desired results.

The Literature Gap

The literature clearly establishes numerous variables in radiology, including spine

imaging, which complicate the interpretive and diagnostic process. This includes the

growing burden of complex data, human bias, individual biological variability,

heterogeneity of pathology, and the multifocal nature of spine pathology. Additional

challenges associated with the use of AI and radiomic methods in spine care include the

intricacy of structures, proximity of anatomic elements, and limited access to relevant

databases and computational disease models. Technological variables include the lack of

standards and wide range of differences between imaging modalities and protocols. The

literature review revealed an absence of gold standards in radiomics. The potential role

and impacts of radiomics is underexplored in all imaging specialties (Oakden-Rayner et

al., 2017). An exhaustive literature search revealed a growing number of research studies

designed to address the role of AI decision support in radiology. The search revealed

some of the challenges associated with imaging interpretation. Many of the published

research studies address narrow applications of AI and therefore do not address the

challenges associated with its adoption, consistent use, and support.

The diagnostic process in spine care is primarily limited to the history, physical

examination, electrodiagnostic testing, and diagnostic imaging. The literature search

established the absence of reliable serum biomarkers for confirming the presence or

progression of spine disorders. The search also confirmed that traditional needle biopsies

are rarely performed on spine pathology. Neurologic deficits associated with a spine

disorder often represent end-stage pathology and a certain amount of permanency is

likely. For the reasons stated spine care providers of all disciplines are highly dependent

on diagnostic imaging reports for decision support at the point of care. Improved delivery

of spine care will subsequently require the use of advanced imaging and related decision

support for the radiologist resulting in more precise and personalized reporting.

The literature search and review performed for this study supported the need to

improve the differential diagnostic process during the interpretive stage of spine imaging

workflow. Image interpretation and reporting in other fields has also been described as

incomplete, inconsistent, and inconclusive (Bosmans et al., 2011; J. Y. Chen et al., 2017).

Wu et al. (2018) discussed the “unmet need for methods that allow more comprehensive

disease characterization and reliable prediction or early assessment of treatment response

and prognosis toward the goal of personalized or precision medicine” (p. 125). Spine

disorders represent one of the most common causes of pain and disability; therefore,

radiologist’s should do what is necessary during the interpretive stage of radiology

workflow to improve the accuracy of the diagnostic process and help ensure successful

delivery of personalized spine care.

AI offers potential solutions for spine imaging, although, little attention has been

paid to its potential. The use of AI in radiology has generally been limited and slow due

to the challenges associated with its development and validation (Aerts, 2016). I was

unable to identify any scholarly research articles that addressed the potential applications

or impacts of the digital (virtual) biopsy in spine care. Knowledge about how to

implement radiomic measures into routine radiology practice is also limited (Vallieres et

al., 2017). The potential role of radiomics needs to be addressed within the context of

spine imaging. The literature reveals many of the challenges associated with its use in

other specialties including lack of standardized imaging protocol, limited access to

annotation training datasets, defined meaningful use applications, clinical utility, and

contouring regions of interest (Lai-Kwon, Siva, Lewin, 2018). The literature review

revealed successful use of AI in many areas of non-spine imaging. Established success,

although limited, has involved the use of natural language processing, radiomics, and

computational diagnostics. The success of AI solutions in radiology requires scalable

applications, clinical utility, and seamless integration into radiology workflow (Court et

al., 2016; Syeda-Mahmood, 2018). There is a gap in the literature surrounding the use

and potential use of AI and AI-supported methods during spine imaging workflow. This

includes the topics of NLP and radiomics.

The role of AI in some fields of radiology such as oncology has advanced more

than in spine imaging. Published research surrounding the role of AI use in other

subspecialty fields of radiology provide the foundation for the discussion of its potential

role in spine care and the design of this research study. The limited research on spine

imaging has been associated primarily with automated identification of normal anatomy

rather than addressing the detection and characterization of disease states. A published

letter in 2016 representing the position of leaders of the American College of Radiology,

a cultural authority in the field of radiology addressed the general gaps in knowledge and

research surrounding the use of AI in radiology. Written to the U.S. Office of Science and

Technology Policy, the authors summarized the gaps in the literature with

acknowledgment of the need for further research to help identify how AI can be used to

access meaningful data, enhance the interpretive phase of radiology workflow, improve

diagnostic accuracy, and reduce errors. This request applies to all areas of radiology

including spine imaging. As noted prior, the results of the literature search confirm the

paucity of studies addressing gaps in knowledge regarding the role of AI in spine

imaging.

Summary

The growing appreciation for the heterogeneity and complexity of pathology has

led to the realization that more precise personalized care is possible with the right

decision support. To achieve this goal, the diagnostic process must be more

comprehensive and classifications of pathology expanded. The standards in radiology

must change. The interpretive approach can no longer be limited to one individual’s

visual assessment of overwhelming volumes of two-dimensional images and the

generation of highly variable qualitative reports. The data are too complex and the stakes

are too high. Automated methods of disease detection and characterization are required to

augment the role of the radiologist. The clinical utility associated with the adoption of AI

solutions in radiology, and, more specifically spine care, must be determined. Timely

personalized patient care should take priority.

An extensive scholarly literature review revealed that AI systems are capable of

integrating and analyzing structured and unstructured data to refine the diagnostic

process. The AI methods required to accomplish this goal include quantitative imaging

with feature analysis (radiomics), acquisition analysis of qualitative imaging features

from records (natural language processing), unsupervised feature learning (text and

images), and scaling of AI solutions to accommodate multiplatform data (distributed

computational models). Successful integration of AI solutions during the interpretive

stage of spine imaging workflow will reduce errors and result in unprecedented

diagnostic capabilities. AI can provide new perspectives of disease that will lead to more

efficient and effective care. Success requires that gatekeepers of big data, such as

radiologists, must accept new responsibilities, assume new roles, and embrace AI

decision support.

This study focused on the potential role and impacts of AI applications during the

interpretive stage of spine imaging. The results of the literature review served as critical

determinants of the potential applications of AI in spine care. The literature review

established the benefits of using AI supported method such as radiomics, natural

language processing, and computational disease modeling in other fields of radiology to

achieve a more precise probability-based diagnosis. It is evident based upon an extensive

review of the literature that in the near future, the interpretive stage of spine imaging will

likely rely on the use of collective intelligence derived from the integration of human and

machine intelligence. This research study was designed to help determine how and when

this might occur.

The published research suggests the radiology workstation will become a hub of

convergent information from other patient diagnostic procedures and databases, including

laboratory, genetic and pathology test results. Databases will include accessible

computational disease models and disease registries. Integrating AI and radiomic

methods will fundamentally alter how disease is diagnosed, classified, and treated (Langs

et al., 2018; Shaikh et al., 2018). Further development of radiomic methods will support

applications for the assessment of non-cancer-related spine pathology.

Current research acknowledges that emerging AI solutions are capable of

providing radiologists with contextually relevant and probability-based differential

diagnostic considerations during the interpretive stage of imaging workflow. This process

improves diagnostic precision and therefore can help overcome human limitations and

bias. In the future, whoever has access to the best data and the best decision support will

likely provide the best care.

Ongoing advances in diagnostic imaging will continue to challenge and expand

our current understanding of disease and related diagnostic criterion. The differential

diagnostic process will soon no longer be limited to the expertise and skills of an

individual; instead, a whole systems process will involve collective intelligence derived

from the contributions of humans and machines. For this outcome to be possible, many

unknown factors must be addressed. This includes what defines meaningful use and

adequate training of an AI system. Heightened awareness of improved accuracy and

efficiency associated with AI decision support will drive computational analytics and

radiomic methods to the forefront of radiology, supporting their eventual role as routine

procedures. Multidisciplinary research will lay the foundation and pave the way for the

transformative process.

Heightened awareness of AI potential and clinical utility in spine imaging is

required to further the research and development process. The endeavor will require the

insights and participation of numerous experts such as AI developers, physicists,

radiologists, pathologists, key influencers, and early adopters of AI in radiology. The

establishment of expert opinions and predictions can help direct further discussion and

research of the topic. A well-designed exploratory case studies can provide this necessary

foundation.

Chapter 3 introduces the research design and methodology used to explore the

potential impacts of AI on the interpretive stage of spine imaging and on the differential

diagnostic process. The results of the extensive literature search reported in this chapter

are used to support the chosen research methodology and strategies acknowledged in

Chapter 3. The chapter addresses my role as a researcher, as well as the data acquisition

and data analysis strategies used in the study. Special attention is placed on the

implementation of steps to protect research participants and to improve the credibility and

trustworthiness of the study.

Chapter 3: Research Method

Introduction

The primary goal of this study was to establish the potential impact of AI

solutions on spine imaging interpretation and diagnosis. I placed special emphasis on the

potential role of radiomics. The unit of study was the interpretive stage of spine imaging

workflow used to detect, characterize, and monitor pathology. The sources of data

included document review, reflective journaling, and focus group sessions. Focus groups

are an effective method for exploring attitudes, expectations, and potential applications

associated with emerging technologies and related processes (Kitzinger, 1995).

In Chapter 3 I introduce the research design and methods used to address the topic

of study. In this discourse I address my role as a researcher and the role of research

participants. I also address the methods used for data acquisition and analysis. The

chapter indicates the research steps implemented to improve the trustworthiness of the

study, as well as the processes used to help ensure the ethical treatment of participants

and the ethical management of data.

Research Design and Rationale

I used a qualitative exploratory case study design to investigate the potential

impacts of AI on the interpretive stage of spine imaging workflow. This approach offered

a flexible and inductive method for acquiring holistic and in-depth insight. The primary

purpose of qualitative exploratory research is to reveal the potential contributions and

influence of an emerging technology or process (Baxter & Jack, 2008). My chosen

research design was used to address how and why questions surrounding the potential use

100

AI applications such as radiomics, NLP, and diagnostic inference methods using deep

learning approaches. Exploratory approaches are often used to lay the foundation for

additional methods of inquiry such as quantitative and mixed method research (Creswell,

2013; Patton, 1990). AI represents a bridging technology comprised of an assemblage of

evolving elements and processes that are sometimes difficult to identify and assess. This

scenario contributes to complexity and uncertainty in research. A qualitative exploratory

approach is able to reveal contextual relationships not adequately addressed by more

restrictive explanatory or quantitative research methods (Ponelis, 2015; Yin, 1984). It

was necessary for me to offer an inductive contextual perspective of potential AI

applications which might be of value to radiologists and other stakeholders in the field.

The chosen study design supported the triangulation of qualitative data acquired

from numerous sources, which helped to improve study validity and trustworthiness. The

qualitative research approach supported purposive sampling, the acquisition of expert

insight from different sources, inductive investigation, and the ability to formulate a

contextual narrative summary. The primary research question was: What are the opinions

of experts regarding the potential use and impact of AI during the interpretive stage of

spine imaging workflow? I added supportive research questions to address determinants

of AI adoption and various applications of AI such as radiomics and natural language

processing.

I acquired qualitative data from expert documents in the form of white papers

published by thought leaders and radiology organizations which addressed the evolution

of AI in radiology. The documents I used represented consensus opinions on the use and

101

potential use of AI. I framed the method of inquiry with insight acquired from an

extensive literature search and from perspectives offered by a consensus-based

prospective document prepared by the Spinecare Data Science Committee of the

American Academy of Spine Physicians (AASP). The committee provided a list of high-

priority topics (needs analysis) related to the potential use of AI during the interpretive

stage of spine imaging. I considered the proposed topics during my development of

research strategies and in my preparation for the focus group sessions.

Qualitative data acquisition occurred in the following sequential stages: an

extensive literature review, review of a prospective document from the AASP Spinecare

Data Science Committee, review of expert documents, and the use of two focus group

sessions, one consisting of radiologists and the other AI experts. I performed reflective

journaling during the entire data acquisition and data analysis process. I used focus

groups to acquire expert knowledge, opinions, perceptions, and predictions relevant to the

topic of study. The research process was designed to reveal themes, noteworthy quotes,

and new perspectives surrounding the use of AI during spine imaging interpretation. Prior

research had established that exploratory focus groups can be used to identify attitudes,

discover opportunities, generate ideas, and frame new questions for future inquiry (Breen,

2006). I used focus group sessions to facilitate creative discussion and to expose the

potential benefits associated with AI use at the radiology workstation in spine care. The

process revealed new insights and exposed culturally formed attitudes and opinions

surrounding the potential applications of AI. The insights I acquired from the focus group

sessions helped me predict the type of synergies, controversies, and debates which may

102

arise surrounding this topic in other research settings. The design of this research study

supported the transition from shared experiences and insights to higher levels of

abstraction and application.

Role of the Researcher

The role of the researcher is important in any research study but particularly

important in qualitative exploratory studies because of the subjective nature of the

process. Researcher experience, motives, and bias can all have a significant impact on the

research process including the analysis and interpretation of acquired data (Durdella,

2019). A qualitative researcher often assumes a primary role in the acquisition and

analysis of data. I assumed the role of the sole researcher in this study. As the sole

researcher, I represented the primary instrument for the collection and analysis of data.

My clinical experience combined with my desire to help identify new forms of decision

support in spine imaging motivated me to pursue the topic of study. It also increased the

risk for professional bias during the research. I subsequently implemented numerous

steps in the research process to reduce my potential for introducing bias into the study.

The steps taken to reduce my personal or professional (researcher) bias and to improve

the trustworthiness of the study included the use of independent experts to review the

focus group moderator guide, the use of member checking (respondent validation), the

application of within group and between group analysis, and triangulation of data. Field

testing of focus group protocols and related research questions was performed to help

establish credibility and the relevance of the approach. I used reflective journaling to help

103

reveal my biases, thoughts, and opinions throughout the research process and provide the

basis for some of the decisions I made during the research process.

Prior to this study I had extensive clinical experience in diagnostic neurology,

neuroradiology, and spine care. I also had extensive academic training in AI and

radiomics. My professional experience as a clinician combined with my familiarity with

AI provided me with the insights required to develop and implement effective

exploratory and analytic strategies. Prior to and during the study I prioritized conducting

myself and the research process in accordance with acceptable scientific methods and in

accordance with the Walden University Institutional Review Board (IRB) guidelines. I

implemented numerous methods to help support scientific data analysis the disclosure of

research conclusions in an unbiased and objective manner. I implemented the previously

disclosed strategies to help ensure that I was reflective and transparent throughout the

research process.

Personal and Professional Relationships

Prior to or during the course of this research study I did not have any formal

business relationship with IBM, any other AI-related company, or professionals who

participated in the focus groups sessions. I did not pursue or accept research participants

with whom I had any prior business relationship. During the proposal stage of the

dissertation process I participated in numerous conference calls with IBM staff including

data scientists to discuss gaps in the research, research strategies, and the management of

research related data. During the early stage of the dissertation process I used a key

contact from IBM to help identify a few renowned AI experts who met the study

104

inclusion and exclusion criteria. I chose to approach IBM due to their performance

record, current market position, and their potential for developing AI solutions for

radiology.

Researcher Bias

Among the many potential sources of bias in a qualitative exploratory research

study, some involve the researcher. Bias can occur in many forms and can influence

different phases of the research process such as the development of research questions,

participant recruitment, expert interviews, data acquisition, and data analysis (Creswell,

2013). Potential sources of bias in this study include the effects of the researcher on the

study and the effects of the research process on the researcher. Researcher bias is possible

whenever research relationships could lead to future business opportunities. For this

reason, I did not accept any proposals or entertain discussions about potential future

relationships.

I was not offered any position with research participants and/or companies or

institutions they were been affiliated with prior to or during the course of the research

study. I pursued the research topic and study with bias in favor of the eventual use of AI

solutions to improve diagnostic accuracy and the interpretive diagnostic process in spine

imaging. I fully disclose that I do not fully understand how AI could or should be used.

As a practicing neurologist, I acknowledge that the accuracy of the diagnostic process in

spine imaging must be improved. Approximately two years ago I sat at a prototypical

IBM Watson AI workstation and experience its potential contribution to the differential

diagnostic process in radiology. With the exception of the isolated experience with IBM

105

Watson, I have no other practical hands-on experience with the use of AI in radiology or

spine care. I am therefore not biased toward the adoption of AI in radiology based on

personal hands-on experience.

Ethical Issues Surrounding the Researcher

As the researcher in the study, I anticipated and addressed potential ethical issues

that may have arisen prior to, during, or after the research process. Prior to designing the

study, I became familiar with the Belmont Report (1979) published by the U.S.

Department of Health, Education, and Welfare, which acknowledged ethical principles

and guidelines which can be used to protect human subjects while conducting research. In

addition, during the course of my PhD studies at Walden University, I completed the

National Institutes of Health (NIH) web-based training program titled “Protection of

Human Research Participants” (Certificate # 2872343). I assumed the duty as the primary

researcher in this study to handle myself in a scholarly fashion and to treat all research

participants in a professional and ethical manner. This required the implementation of

steps to ensure participants well-being while protecting their rights and minimizing their

exposure to potential harm.

Research Methods

The dissertation proposal was accepted and Walden University IRB approval was

obtained prior to beginning the research process. The IRB reviewed the research plan and

the research methodology, as well as all pertinent documents. The approval process was

completed prior to research participant recruitment and the acquisition of data. I

implemented steps to disclose and address researcher bias, potential conflicts of interest,

106

competing motives, and potentially detrimental power relationships that could have

developed during the course of the study.

Study Population and Sampling

The study population consisted of stakeholders involved in or influenced by the

interpretation of spine imaging. This includes data scientists, device manufacturers, AI

programmers, radiologists, spine care providers, and other health care providers. It was

important to identify the subpopulation of stakeholders most capable of addressing the

exploratory research topic within focus group settings. The success of this study

depended on my ability to recruit research participants experienced and knowledgeable

on topics related to the potential impact of AI used during the interpretive stage of spine

imaging workflow. The focus group study population was limited to radiologists and AI

experts. The members of each category of participants were intricately involved in

processes which took place within the parameters of the unit of study, which was the

interpretive stage of spine imaging. I implemented steps that required that all research

participants met strict research inclusion and exclusion criteria.

A certain degree of homogeneity or similarity within a focus group session

combined with purposive sampling of professional participants has proven to enhance the

potential for exploring new technology (Kitzinger, 1995). Kitzinger (1995) also

demonstrated that a group of professionals with common knowledge along with similar

training and experience are more likely to engage in in-depth discussions. To facilitate

this approach, I placed AI experts in one focus group session and radiologists in a distinct

and separate focus group.

107

Population sampling for the research study was convenient and purposive so that

small groups of confirmed experts could discuss and explore the research topic. Experts

recommend purposive sampling strategy to help ensure that participants have the level of

experience and expertise required to contribute to the topic of study in group sessions

(Creswell, 2012; Patton, 2002). Subsequently, I used purposive sampling in this study to

help ensure that all of the research participants had an adequate level of expertise,

interest, and experience surrounding the development or application of AI solutions

during the interpretive stage of diagnostic imaging. The use of convenience sampling

combined with purposeful sampling brought together like-minded professionals who

were familiar with the topic of study and categorically with role of AI experts and

radiologists in spine care.

The population sampling strategy included estimation of the sample size required

to achieve the level of representation and topic saturation required to explore the potential

impact of AI on spine imaging interpretation and diagnosis. A renowned key AI contact

was used to help identify qualified AI professionals to participate in the study. I used key

radiology contacts to help identify radiologists qualified for participation in this study. I

provided each of my contacts with the purpose of the research study, as well as

participant inclusion and exclusion participation criterion prior to asking for their

recommendations. I confirmed that all potential participants met inclusion and exclusion

criterion prior to their acceptance into the study.

The ideal size of a focus group is often five to eight participants, unless more are

required to address a complex topic (Krueger & Casey, 2015). Qualitative research

108

experts have acknowledged that four to 10 individuals are often adequate for

homogenous sampling of professionals (Creswell, 2013; Krueger & Casey, 2015). I

limited the number of expert participants to eight to 12, or four to six in each of the two

focus group sessions, to facilitate creative and in-depth discussion of a complex and

contemporary topic. The small focus group sizes helped me as the moderator better

manage research topics and the flow of discussion and ensure that each participant had

adequate opportunities to share their expertise and insights. The use of two homogenous

focus groups comprised of confirmed experts was large enough for this study to provide a

diversity of perspectives and opinions.

The research participants accepted for participation in this study were from

different clinical and professional settings. I recruited research participants through

personal contacts. I followed up with potential participants by of email and phone calls. I

did not offer material or monetary incentives during the recruitment process. Research

subjects who agreed to participate in a focus group session were asked to complete a brief

survey prior to the focus group session (Appendix B). I used a survey to acquire

participant demographic information such as their current position, background, and

experience.

Prior to convening for focus group sessions each research participant received an

acceptance letter which included an introduction to the research project, a focus group

agenda, a list of what was expected of them and a research participation consent form to

review, sign and return. Each research participant received email notification of the

scheduled focus group session along with an invitation to participate in person or via a

109

prearranged teleconference option. I arranged for the teleconference option through

Zoom, a highly respected and secure service. Email notifications included the focus

group facility address, focus group directives, and access information for the

teleconference option. Each participant received periodic email reminders of their

scheduled focus group session. I requested that research participants register online in

advance of the focus group sessions.

Sample size in qualitative research is often not as important as the chosen

methods of data acquisition, data analysis, and data validation (Njie & Asimiran, 2014).

The potential use of AI during the interpretive stage of spine imaging is both a new and

complex topic. Subsequently, I chose small sample sizes to help achieve expert in-depth

discussions and topic saturation during the focus group sessions. Data saturation is

reached in a qualitative study when a coherent and consistent perspective is reached

(Guest, Bounce, & Johnson, 2006). I established the criteria for data saturation in this

study prior to the focus group sessions and included the definition on in the moderator

guide (Appendix D). I developed open-ended and probing research questions to help

achieve data saturation during the focus group sessions.

Research Participant Inclusion and Exclusion Criteria

In qualitative research, research participants must be capable of contributing to

the study with the chosen methods of inquiry (Creswell, 2013), particularly when

addressing a complex and rapidly evolving technology such as AI. I used convenient and

purposeful sampling methods to select research participants for this study. The potential

research participants were subjected to explicit study exclusion and inclusion criteria. The

110

inclusion criterion addressed the potential participant’s current professional status,

background, and experience. The inclusion criterion for AI expert participation was a

minimum of five years of experience in health care AI development or applications. In

addition, each AI participant was required to have a minimum of a bachelor’s degree in a

related field such as informatics, data science, computer science, or AI. AI participants

were also required to be actively working in an AI field.

The inclusion criteria for radiologists were a minimum of 10 years of experience

in spine imaging interpretation. In addition, each participant was required to hold a

doctoral degree and to be actively working in a diagnostic radiology capacity. Each

radiologist was required to be board-certified in a field related to the topic of the study.

Study exclusion criteria for the AI expert and the radiologist included a history of or

current employment with the Chicago Neuroscience Institute (CNI) or the American

Academy of Spine Physicians (AASP), both of which I am affiliated with. I prohibited

key contacts for research participant recruitment from participating in the study. I also

prohibited professionals who played a role in field testing of focus group questions and

strategies from participation in the research study.

Data Collection Instruments and Processes

The use of predetermined or validated data collection instruments can help direct

the data acquisition and data management processes. I acquired and developed numerous

data collection instruments for use in this research study. This included the use of

qualitative research software, a brief qualitative survey, and a focus group moderator

guide. I developed a brief survey and gave it to each participant to complete prior to

111

participation in a focus group session (Appendix B). This document and the consent form

were used to confirm that the various experts met the criterion for participation. I

developed a focus group moderator guide to help manage time, topic discussions, and the

method of inquiry during the focus group sessions (Appendix F). Published research has

demonstrated that the use of a focus group moderator guide helps ensure efficient and

systematic in-depth coverage of research topics and related questions (Fraenkel &

Wallen, 2003). The moderator guide I used in this study consisted of carefully crafted

open-ended questions and probing semistructured questions.

I developed questions for the focus groups with the assistance of insight acquired

from an extensive literature search and from the needs analysis document provided by the

AASP Spinecare Data Science Committee (Appendix F). Each question was placed in the

focus group moderator guide. My development of the moderator guide was overseen by

independent expert prior to and after field testing. I made all necessary changes to the

guide. This iterative process of assessment helped me reduce the risk for an inappropriate

or biased approach to research question development and delivery. It also helped me

refine the methods of inquiry I used during each focus group session. The focus group

moderator guide consisted of an agenda, an introduction, along with a list of PowerPoint

concept slides, and research questions followed by closing remarks. (Appendix D). The

moderator guide identified the order of topic presentation and inquiry. I used the guide to

help set the tone for each focus group session and for guiding the order of the process.

Consistent with the recommendations of C. L. Lee et al. (2015) I developed data

coding guidelines to help ensure analytical and categorical consistency during content

112

and thematic analysis (Appendix C). My coding guidelines consisted of predetermined

codes capable of being adapted, modified or replaced during data analysis. I developed a

priori codes consistent with the conceptual framework of the study utilizing theoretical

perspectives from DOI and TAM. The a priori codes aligned with the research topic,

research purpose, and research questions. During the course of the entire research

process, I made regular entries in a reflective journal to memorialize my biases,

impressions, and insights.

I used numerous expert documents (white papers) to help identify current

consensus-based opinions and positions surrounding the use of AI in radiology and

oncology. At the time of this study, there were no white papers published on the potential

role or impacts of AI in spine care. I used a few published papers, which addressed

narrow applications of AI in spine care. The consensus-based “white papers” I used for

this study were published by nationally and internationally recognized organizations: the

Canadian Association of Radiology, the American College of Radiology, the French

Radiology Community, and the European Society of Radiology. I also used seminal

publications of leading experts in the field. I analyzed, thematically coded, and

triangulated the content of the expert documents with data from other sources to improve

the consistency and relevancy of the study’s conclusions. This methodical and transparent

analysis process helped improve the trustworthiness and validity of the study. The

document provided by the AASP Spinecare Data Science Committee offered a list of

potentially meaningful applications of AI during the interpretative stage of spine imaging

workflow. The AASP document was developed by a multidisciplinary group of spine

113

care experts independent of the research process. I used this document along with insight

acquired from an extensive literature search to guide the development of research

questions I used in the focus group sessions.

The research process was not be limited by fixed guidelines or rules, subsequently

allowing for inductive assessment of emerging topics and trends. As stated previously,

focus group research represents a well-established and disciplined scientific method for

acquiring in-depth insight surrounding the use of new technology and related processes

(Krueger & Casey, 2015). I applied the concept of data saturation during focus group

sessions and during thematic data analysis of expert documents. I used an introductory

PowerPoint slide program at the beginning of each focus group session (Appendix E). I

conducted topic-specific discussions during the focus group sessions until reasonable

topic saturation was achieved. I arranged for a recording of all of the contributions during

each focus group session. I also arranged for verbatim transcription of the recorded

sessions to avoid misinterpretation or misrepresentation. I performed document analysis

until I achieved topic saturation. I triangulated the data from the different research

sources to improve the internal, as well as external validity of the study. A concise and

comprehensive informed consent form was developed and used to protect the rights of

participants and to encourage unfettered contribution to the research process.

Document review and analysis offers a unique and often critical contribution to

qualitative exploratory research (Creswell, 2013; Patton, 1999). I used position papers

and consensus-based summaries published by reputable organizations and highly

regarded experts in this study to help address the potential impact of AI during spine

114

imaging workflow. I compared and contrasted the data acquired through expert

documents with data acquired through other expert sources such as the focus group

sessions and reflective journaling to improve the internal validity and transferability of

the results.

Data Acquisition

I initiated the data collection and analysis process after I received dissertation

proposal approval and Institutional Review Board (IRB) approval from Walden

University. I created a data acquisition flow diagram to help guide me in the research

process. The method of inquiry I used in the focus group session was field tested with

two independent experts, one meeting AI expert participation criteria and the other

radiologist criteria. I did not accept the experts who assisted me with field testing as

research participants. I used field testing to evaluate the focus group protocols and

strategies I used in the Focus Group Moderators Guide. I made minimal modifications, as

a result of the field testing.

The focus group sessions each lasted approximately 90 minutes. I achieved an

acceptable degree of data and topic saturation in each session. I arranged for each focus

group session to be digitally recorded, transcribed verbatim, and stored securely.

Emotional responses, body language, and nonverbal forms of communication between

the participants in a focus group setting can be important (Bunnick et al., 2017). I

subsequently recorded any participant behavior during the focus group sessions I felt was

relevant to the study purpose. I led each focus group session with the assistance of the

moderator guide. I developed my focus group approach guided by Krueger’s categorical

115

strategies that included the use of an opening question, introductory questions,

transitional questions, and probing questions (Krueger, 2000). I used probing questions to

help actively engage each of the participants in topic discussions.

Scholarly publications have established that reflective journaling offers the

researcher with a powerful inductive method for recording observations ideas, and

insights using an active voice (Janesick, 2011). I performed reflective journaling during

the course of the research process. Journaling served many purposes. It allowed me to

identify my initial and evolving perspectives, expectations, and biases associated with the

research topic and the research process. Review of journal entries gave me the

opportunity to engage a higher level of critical thinking and implement methods to reduce

my personal influence on the research process and outcome. Journaling included my

impressions of verbal, as well as nonverbal communication during focus group sessions.

This included recoding of body language and expressions. I also recorded the level and

nature of agreements or disagreements that occurred during each focus group session.

Research Participant Debriefing and Follow-Up

At the completion of each focus group session, I reminded participants of the

purpose of the research study and informed them how I would manage and analyze the

data acquired. In informed the research participants that they would receive an overview

of the focus group data analysis in the form of a thematic summary and a list of

supportive quotes for review, a process referred to as member checking or respondent

validation. In addition, I informed each research participant that the records of the

research study including their consent forms would be stored in a secure location for a

116

minimum of 5 years, after which time they would be properly destroyed. I informed each

participant they would receive notice when the dissertation was published. I also assured

each research participants that they would be provided with access to the published work

when it was available.

Data Analysis

I used a multistep process to analyze acquired data. I implemented an inductive

and iterative data analysis process as soon as data was acquired. My analysis process

continued throughout the entire research study. I recorded a chain of evidence to

memorialize the process and analyzed the focus group transcripts with an exhaustive,

inductive, and iterative process of coding for themes. Descriptive codes were clearly

established and defined consistent with the work of Glaser and Laudel (2013). I used a

hybrid approach to coding, allowing for aggregation, subtraction, combining, and

expanding of code categories when necessary. Qualitative data coding offers an effective

method for revealing emergent ideas, themes, and relationships (Rubin & Rubin, 1995;

Strauss & Corbin, 1998). My evaluation of the focus group transcripts included content

analysis and thematic coding. Content analysis is used in the social sciences and in

qualitative research to structure information (Krippendorf, 2004). The analysis of focus

group data should include identification of noteworthy quotes, as well as identification of

outlying factors and unexpected findings consistent with published works (Breen, 2006).

I identified and labeled all noteworthy findings, comments, and quotes derived from my

assessment of the various sources of research data. I performed the data coding process

until topic saturation was achieved.

117

The first step in the coding process was to become familiar with the data and

systematically reduce its complexity. I developed a few provisional (a priori) codes to

initiate axial coding. I developed the provisional codes with insights acquired from my

extensive literature search along with the influence of theoretical constructs from DOI

and TAM. Some of the provisional codes aligned with theoretical constructs of DOI and

TAM, such as relative advantage, interoperability, complexity, ease-of-use, and perceived

usefulness. I replaced, revised or modified many of the provisional codes during data

analysis to better describe and label acquired data. I expanded, contracted, replaced, and

modified the coding categories many times throughout my analysis process. Thematic

coding arose from the integrated applications of provisional coding, open coding, in vivo

coding, axial coding, and selective coding.

My analyses of focus group data included within and between group analysis. I

took into account the unique experiences and backgrounds of the participants in each

focus group session revealed by their demographic surveys. I displayed the results of my

analysis of the acquired research data in many different ways, including a contrast table.

Contrast tables offer an effective method for looking at relationships between exemplars,

extremes, and outliers (Miles, Huberman, & Saldana, 2014). Combining content analysis

and thematic coding supported the development of a concept map depicting the potential

relationships between processes and technologies, used during the interpretive stage of

radiology workflow. I created a concept map to reveal relationships between the flow of

data and processes during the interpretive stage of radiology workflow. I used the concept

118

map to help transform tacit knowledge into a practical resource and a foundation for

further discussion and research.

Concept Mapping

Data analysis in qualitative research involves many steps that include data

reduction, data organization, data interpretation, and data display. Concept mapping has

been successfully used to graphically organize and depict relationships between elements

of a system or process (Baugh, McNallen, & Frazelle, 2014). This includes the

relationships between data, individuals, and technology (Baugh et al., 2014). Concept

mapping has also been used in qualitative research to reveal themes and to depict

workflow (Daley, 2004; Novak, 1998). One of my goals in this research study was to

identify themes that could be used to create one or more concept maps depicting the

potential role of AI solutions during the interpretive stage of spine imaging workflow. A

concept map helps depict stages of a complex process and reveal technological

relationships to achieve desired goals (Daley, 2004; Novak, 1998; Wheeldon & Faubert,

2009). I subsequently developed a concept map to reveal the flow of data and role of

potential AI applications during the differential diagnostic process associated with

interpreting spine images.

I used concept mapping in this study to facilitate a shared vision, to help direct

subsequent research, and to inform further technology development. I also used it to help

determine how to embed AI technology into existing radiology workflow. A concept map

can be augmented with the use of numerous elements such as linked tasks, labeled

processes, communication pathways, and hierarchies of priority (Wheeldon & Faubert,

119

2009). In my concept map I used map lines to link mapped elements and I used

directional arrows to reflect the flow of data and/or the implementation of a process. I

developed a concept map in this study to complement and enhance textual conclusions.

This helped me present research findings in an accurate, concise, and effective manner.

Trustworthiness of the Study

The trustworthiness of a research process is influenced by the role of the

researcher, the source of the data, the management of the data, and the approaches used to

improve study validity and reproducibility (Connelly, 2016; Mays & Pope, 2000;

Shenton, 2004). Qualitative exploratory case study research is inductive and subjective

and therefore requires high levels of trustworthiness, reliability, and validity to be

influential (Creswell, 2013). The attributes of validity and reliability are operationalized

in different ways in qualitative versus quantitative research studies (Mays & Pope, 2000).

The primary risks associated with qualitative case study research include over

generalization of results, researcher bias, inadequate interpretation of data, poor

integration of data, and research question mismatch with methodology. I took extra

precautions and implemented steps throughout this research process to improve the

trustworthiness of the study and its conclusions (Figure 5).

Qualitative studies that include the use of focus group sessions must meet

extremely high standards to be reliable and valid. Research study trustworthiness is

determined by its credibility, confirmability, dependability, and transferability (Lincoln &

Guba, 1985; Shenton, 2004). I addressed each of these elements in this study along with

120

reliability. I implemented numerous steps d to reduce the risk for interjecting personal

bias and to support evidence-based conclusions.

Credibility

I implemented numerous steps to improve the credibility of this research study. I

used respondent validation also referred to as member checking to help confirm the

accuracy of my thematic conclusions. I also provided research participants with a list of

the supportive quotes I acquired from the focus group transcripts. Member checking is an

important step in qualitative research, because it provides participants with an

opportunity to affirm the accuracy of focus group data acquisition, analysis, and

interpretation (Creswell, 2013). Member checking in this study served as an effective

method for establishing interpretive and descriptive validity. It also helped reduce the

impact of my personal bias as the sole researcher.

To help further reduce personal bias during data acquisition and analysis, each

focus group session was recorded and transcribed verbatim. I used a few open-ended

questions during the focus group discussions to help reduce the risk of framing bias. I

read the focus group transcripts numerous times to ensure comprehension of the material

prior to initiating descriptive labeling and coding of data. I used an inductive and iterative

process of hierarchical coding to avoid rigid misclassification of data.

I used a well-defined unit of analysis to help direct the research process and the

flow of data. This approach improved study credibility. Consistent with the work of Mays

and Pope (2000) I performed reflective journaling to expose how my role as the sole

researcher may have influenced the research process and outcomes. I used reflective

121

journaling to record my thoughts and opinions during the entire research process. The

journaling process helped reveal my perspective and beliefs and how they may have had

an impact on data analysis, research design, and research conclusions.

Triangulation of data acquired from different sources improved the credibility and

internal validity of this research study. Published work has demonstrated that the

triangulation of data acquired from a diverse set of expert sources contributes to the

authenticity, plausibility, and validity of qualitative research (Greenlaugh & Singlehurst,

2011). Triangulation of acquired data from the focus group sessions, reflective

journaling, and from consensus-based white papers in this study improved the credibility

of the research conclusions. My use of theoretical constructs from DOI and TAM

combined with insights acquired from an exhaustive literature search helped reduce

personal bias during my formulation of research questions and the interpretation of the

responses. In summary, the methods I used to improve the internal validity of this study

included reflective journaling, respondent validation (member checking), inductive

coding, and triangulation of data from diverse sources.

Transferability

Transferability refers to the ability of a reader to apply the research process or the

research results to another situation or setting. The success of this process is dependent

on transparency and adequate description of research boundaries, parameters, and

processes (Connelly, 2016; Lincoln & Guba, 1985; Shenton, 2004). In contrast to

transferability, generalizability refers to the ability to apply the results from a research

sample to a broader population. In order for research to be transferable, the results must

122

be reproducible in different cultural settings with common variables (Shenton, 2004). In

this study, I used sequential steps, thick descriptions, a data acquisition flow chart, a

focus group moderator guide, and qualitative coding software, all of which contributed to

a high degree of transparency, which supports transferability.

I used numerous redundant and overlapping methods to improve external study

validity and transferability. I disclosed unexpected and conflicting results along with

unforeseen challenges in the research. My research conclusions include alternative and

rival explanations surrounding the potential impact of AI use during the interpretive stage

of spine imaging workflow. The use of two focus group sessions each comprised of

homogenous groups of experts from two related fields supported the detection of patterns

and themes within and across groups. I compared the findings of the focus group sessions

to the themes that emerged from consensus-based white papers, which served as research

documents in this study.

Dependability and Confirmability

The attributes of dependability and confirmability are important elements of

trustworthiness that influence the ability to replicate the research process and to test

related assumptions. The dependability of qualitative research improves with transparent

strategies such as thematic coding, content analysis, and the generation of thick

descriptions (Shenton, 2004). My use and disclosure of a field-tested focus group

moderator guide and data-coding guide offered the level of transparency required to

facilitate accurate interpretation and/or replication of this research. I used valid tools and

measures available on the Atlas.ti, Version 8 software, to analyze and manage the

123

research data in this study. I made journal entries of influential issues surrounding the

integrity or quality of data used for analysis.

I used clear and concise descriptions of the research process and the flow of data

to improve the ability for others to critique or replicate the study. I also recorded the

chain of evidence and performed content analysis and thematic coding, which I

acknowledged in detail with the help of a hierarchical coding table. I developed a data-

coding guide that includes a list of code categories and their definitions and the criteria

and methods I used to achieve and define data saturation during the analysis process. I

performed within and between case analyses with the help of computational methods to

reduce my potential bias.

Ethical Procedures

The basic ethical principles acknowledged by the Belmont Report (1979) are

respect for individuals, beneficence, and justice. Beneficence refers to the treatment of

individuals in an ethical manner by respecting their decisions, securing their safety,

prioritizing their well-being, and protecting them from harm. Justice refers to the equal

and fair management of research participants. To confirm adherence to these basic

principles I treated each research participant equally. I provided each research participant

with the same documents, had them sign the same consent forms, and exposed them to

the same data collection processes. The research protocol and conduct in this study

conformed to the tenants of the Belmont Report and to Walden University IRB

requirements.

124

Documents and Agreements

I performed the research in an ethical and honest manner to ensure the integrity of

the study and to minimize any potential harm or risk to research participants. I used

predetermined protocols and preapproved documents helped to ensure that an appropriate

and ethical approach was used throughout the entire research process. Research in the

Walden University doctoral program requires oversight by the IRB to help ensure the

integrity of the research process, as well as the safety and privacy of all research

participants. The Walden University IRB approval number assigned to this study was 02-

13-19-0129405.

Prior to collecting the data, I provided each research participant with preapproved

documents, which included an overview of the study, a focus group agenda, participant

expectations, and an informed consent form, which included confidentiality terms. The

documents safeguarded the consistent and ethical treatment of each research participant

and the ethical management of research data. The agreements disclosed any anticipated

or potential exposure to risk. The documents also acknowledged the voluntary nature of

study participation.

All of the research participants were required to sign an IRB approved consent

form prior to participating in the study. The form included confidentiality agreements.

Consent included my responsibility as the researcher to keep confidential the personal

identities and contributions of all research participants. I informed all participants of the

purpose and scope of the study, as well as the expectations for their participation. In

125

addition, I informed the research participants that they would each receive a nominal

stipend of $25 for participating in the study.

Treatment of Research Participants

I treated all of the research participants with the utmost respect. Research

participants should also be treated as autonomous agents (Kaiser, 2009). The safety and

rights of research participants must be prioritized at all times (Belmont, 1979). In

qualitative research, the researcher assumes a unique responsibility for protecting each

research participant (Orb, Eisenhauer, & Wynaden, 2001). Compliance with well-

established ethical principles such as autonomy, beneficence, and justice helps ensure

proper care of research participants (Lorell et al., 2015; Orb et al., 2001). Consistent with

the recommendations of Kaiser (2009), the methods and forms I used to obtain informed

consent in this study were adapted to the type of research and type of research

participants required. My use of field testing and an independent review of the focus

group moderator guide helped guarantee appropriate treatment of research participants in

this study.

I informed all of the research participants of their right to withdraw from the

research study at any time and for any reason. In addition, I provided participants with

the option to withdraw verbally or in writing from the study without any repercussions. I

informed all of the research participants that their names would be kept confidential and

their identities would remain anonymous. I removed research participant names from

focus group transcripts and replaced them with unique and anonymous identifiers to help

126

ensure confidentiality of their names and contributions. Examples include Participant 1

(P1), Participant 2 (P2), and so forth.

Treatment of Data

To help safeguard trust and to protect the privacy and contributions of each

research participant during the focus group sessions, participants were informed that they

are not to disclose the names or contrition of participants outside the research setting. I

deleted all names from the final focus group transcripts. The research records will be

securely stored for 5 years from the time of completion of the research study to protect

the rights of all participants. Proper storage of research records will support authorized

access for auditing or for review by qualified individuals. I will take proper steps to

discard all participant records after 5 years.

The Potential for Research Impact on Social Change

The primary purpose of this research study is to explore the potential impacts of

AI on the interpretive stage of spine imaging and to reveal its social implications. The

study addressed the potential influence on standards of care, technology development,

systems applications, and public expectations. In Chapter 5, I expand the discussion of

the research results to include it potential impacts on various levels of society.

Meaningful use of AI during the interpretive stage of spine imaging will augment

the role of the radiologist by reducing data complexity, characterizing pathology, and

offering decision support. The process will empower the radiologist as a gatekeeper of big

data and facilitate their leadership role. Improved availability of AI decision support will

increase the demand for remote access teleradiology services. The process has the

127

potential to support improved democratization of decision support surrounding diagnostic

image interpretation. It will subsequently offer a potential solution for underserved

professionals, facilities, institutions, and geographic locations.

Widespread use of AI use during radiology workflow may alter the health care

landscape, especially in the areas of image analysis, disease characterization, disease

monitoring, decision support, and final report generation. AI could offer radiologists new

solutions, capable of improving their ability to detect early stage pathology. Most

diseases are recognized at advanced stages, thus, resulting in high costs and poor

treatment outcomes. Early disease detection would contribute to more efficient care at

lower costs. These outcomes would all have a favorable social impact at many levels.

Successful use of AI during the interpretive stage of radiology workflow will

require redesign and co-evolution of supportive technologies to benefit the broader field of

health care and society. Supportive solutions will include new levels of interoperability

between databases and management systems (Tang et al., 2018). A successful co-

evolutionary process will require the development of unifying platforms, which facilitate

sharing of data and support more consistent use of disease criteria, disease classifications,

and computational disease models. The summary discussion addresses the potential

relationship between AI and relevant emerging technologies. For example, block chain

technology has the potential to provide proof-of-work validation while recording the flow

of data and computational steps across an AI-based network (Kuo, Kim, & Ohno-

Machado, 2017; Mamoshina et al., 2018). The eventual convergence of AI and block

128

chain technology has the potential to decentralize intelligent decision support and offer

increased access to computational disease models.

In summary, AI decision support could overcome human bias, reduce interpretive

error, and enhance the potential for a more precise and timely diagnosis in spine imaging.

Meaningful use of AI during the interpretive stage of spine imaging and at other levels of

radiology workflow could have a favorable impact on the role of radiologists, as well as

on the co-evolution of decision support technology, standards of care, delivery of care, and

public expectations. I address the potential social consequences of AI development and

use in spine imaging in the conclusion of this study.

Summary

I designed this research study to identify how the use of various AI solutions

could impact data management and the differential diagnostic process during the

interpretive stage of spine imaging. AI involves a rapidly evolving set of technologies

associated with numerous processes. Its role in radiology is difficult to define because of

its wide range of potential applications. It was necessary to address the potential impact

of AI with a qualitative exploratory case study approach to identify possibilities worthy

of further discussion and investigation. The design of this study supported the acquisition

of expert insights and data from different sources. This research design also supports the

development of a concept map representing potential AI applications and contributions

during the interpretive stage of spine imaging workflow. The research design and

methods I detailed in this chapter were supported by expert sources, an extensive

129

literature search, and a review of consensus-based documents surrounding the use of AI

in radiology.

An individual radiologist can no longer be required to function with precision

accuracy in the face of overwhelming, high velocity, and complex data. Radiologists

require new data analysis and decision support systems during the interpretive stage of

imaging in all fields including spine care. Successful use of AI will likely result in earlier

disease detection, better disease characterization, less diagnostic errors, and shorter

lengths of care (Kohn et al., 2014; Lee, 2017).

The research methods introduced in this chapter provide a trustworthy approach

and a scholarly foundation for further discussion and research surrounding the

development and use of AI during the interpretive stage of spine imaging. I designed this

research study to introduce the role of radiomics and the concept of the digital (virtual)

biopsy, in a manner that could be applied in spine care, as well as in other fields of health

care.

There is a growing demand for health care to become more predictive and

preemptive. Success requires a more deliberate approach to the comprehensive and

objective analysis of actionable data. My acquisition and analysis of data in this research

study revealed concepts and themes that can be used to develop additional research

strategies to pursue the role of AI solutions during the interpretive stage of spine imaging

workflow. Discoveries associated with this research can be applied to other areas of

radiology.

130

Chapter 4 presents the results of the research study. I further discuss how I

acquired and analyzed the data and how I improved the trustworthiness of the study and

related data.

131

Chapter 4: Results

Introduction

The primary purpose of this qualitative, exploratory case study was to explore the

potential impacts of artificial intelligence on spine imaging interpretation and diagnosis. I

designed the study to acquire and analyze expert opinions from different sources. My

goals for the study included identifying how AI solutions might improve the accuracy

and efficiency of interpretive workflow and the differential diagnosis process in spine

imaging. I implemented qualitative research methods to explore the possibilities

associated with computational decision support and to establish a thematic basis for

further discussion and research on the topic.

This study is one of the first to address the potential role of AI in spine care and

the concept of the digital (virtual) biopsy characterized by multiscale in vivo

interrogation of pathology. I initiated this study with the fundamental belief that images

are rich in metadata and that diagnostic imaging represents a core diagnostic process in

spine care. Select constructs of the TAM and DOI were used to guide the process of data

acquisition and analysis. During focus group sessions, I asked open-ended and probing

questions to address radiomics, interpretive workflow, the differential diagnostic process,

clinical utility, and determinants of AI adoption and use.

The volume and complexity of data acquired with advanced diagnostic imaging

methods has created a burden and exposed unprecedented opportunities for radiologists.

Big data has exceeded the ability of a radiologist to make fully informed decisions (Aerts,

2017; Gilles et al., 2016). Radiologists require augmentation of their role to reduce errors

132

and to take advantage of new opportunities for rendering a more precise and personalized

diagnosis. Without adequate technological assistance, the human interpretive process

within radiology workflow will become progressively more inaccurate, inefficient, and

untimely (Croskerry 2013; Manrai et al., 2014; Obermeyer & Emanuel, 2016; Ragupathi

& Ragupathi, 2014). Diagnostic imaging represents one of the single most important

methods for detecting and characterizing pathology in all biological systems including

the spine. This study addresses the potential for AI to reveal actionable data from

imaging studies while augmenting the role of the radiologist. My primary motivation for

performing this exploratory study was to acquire insight and provide direction for the

development of decision support solutions to support better spine care.

In this chapter, I address numerous topics such as the research purpose, the

research setting, field testing, research participant demographics, data collection, data

analysis, trustworthiness of the study, and the research results. I laid out this chapter in a

manner consistent with the chronological stages of the research process. In this chapter, I

reveal the themes and subthemes which emerged from triangulation of data and data

analysis. I also provide supportive evidence for the iterative process. This chapter

indicates how the research design and the use of strategic methods improved the

trustworthiness of the research results. In addition, I provide an overview of my reflective

journaling, which includes disclosure of its impact on the research process and results.

The conclusion provides a summary of the research findings along with a transition to

Chapter 5.

133

Field Testing

I field tested the moderator guide to help establish the required level of

appropriateness, clarity, and relevance of the approach used during focus group sessions.

I achieved these goals through independent review of the strategies and resources

outlined in the focus group moderator guide, including the agenda, topic introduction,

ground rules for participation, open-ended questions, and a script for the conclusion of

the session. I also field tested the appropriateness and clarity of my introductory

PowerPoint slides I used during focus group sessions.

I performed the field testing, on separate occasions, with one radiologist and one

AI expert. The experts who participated in the testing met study inclusion criteria but did

not serve as research participants. The field testing process allowed for peer-review of the

focus group protocols and resources. The experts who participated in the field testing

were not asked to answer or respond to any research questions. I did not have an

employment or consulting relationship with the field testing experts.

I used field-testing to help establish appropriate and relevant focus group

protocol. The process included assessment of the methods of data acquisition and the

pattern of inquiry. I made no significant changes as a result of field testing, with the

exception of the order and clustering of questions to be used during the focus group

sessions. Nor did I make contextual revisions to the primary focus group questions. I field

tested which PowerPoint concept slides were the most neutral and concise to help guide

focus group discussions on complex topics. I removed a few slides from the presentation.

I made no content changes to the remaining PowerPoint slides.

134

My field testing helped identify the order of the focus groups. I determined that

the radiology focus group session should take place first, followed by the AI expert focus

group session. The independent experts believed this order would support more

progressive, in-depth coverage of the research topic. I thought it was necessary to utilize

field testing to reduce bias, ensure professionalism, and to efficiently operationalize

available multimedia and data acquisition methods.

Research Setting

I held the focus group sessions at an independent and professional location. The

setting supported physical participation and teleconference access for all participants. I

provided each of the research participants the option to be physically present during the

focus group session or to access the session using Zoom, an independent well-established

teleconferencing solution. Each research participant chose to access the focus group

session with the teleconferencing solution.

During the live focus group sessions, each participant using Zoom had access to

an online image gallery for intimate real-time viewing and interaction with all

participants. Each participant also had the independent option of engaging a speaker

highlight function that prioritized the participant actively contributing. I provided each

participant with access to a dynamic online gallery to facilitate efficient communication. I

served as the sole moderator for each focus group session. I moderated each session from

a conference room designed for focus groups. The room consisted of a boardroom table

and chairs, professional audio system, and a large format wall-mounted screen with a

camera.

Students also viewed