Paraphrasing a paragraph

Citation: Deep learning-based

Ultrasound IMC Segmentation and

cIMT Measurement . Appl. Sci. 2022,

1, 0. https://doi.org/

Received:

Accepted:

Published:

Publisher’s Note: MDPI stays neutral

with regard to jurisdictional claims in

published maps and institutional affil-

iations.

Copyright: © 2022 by the authors.

Licensee MDPI, Basel, Switzerland.

This article is an open access article

distributed under the terms and

conditions of the Creative Commons

Attribution (CC BY) license (https://

creativecommons.org/licenses/by/

4.0/).

applied sciences 1

2

Article 3

Deep learning-based Ultrasound IMC Segmentation and cIMT 4

Measurement 5

Hanadi Hassen Mohammed 1* , Omar Elharrouss 1, Najmath Ottakath1, Somaya Al-Maadeed 1, Muhammad E. H. 6

Chowdhury 3, Ahmed Bouridane 3, and Susu zughaier4 7

1 Department of Computer Science and Engineering, Qatar University, Doha, Qatar 8 2 Department of Electrical Engineering, Qatar University, Doha, Qatar 9 3 Cybersecurity and Data Analytics Research Center, University of Sharjah, Sharjah, Unit-ed Arab Emirates 10

Abstract: Common Carotid Intima-Media Thickness (CIMT) is a typical atherosclerotic measure 11

frequently estimated in carotid ultrasound images. The use of deep learning as a promising method 12

for medical image analysis, segmentation, and CIMT measurement is lacking in such kinds of images. 13

In this paper, the CUBS dataset (acquired from both sides of the neck of 1088 participants) is used for 14

evaluating the intima-media complex (IMC) segmentation performance for four recent deep learning 15

models, namely, CNN-based Network, Self-ONN-based Network, Transformers-based Network, and 16

Pixel Difference Convolution-based network. The analysis of the segmentation performance shows 17

that the Self-ONN-based model outperforms the conventional CNN-based model while the Pixel 18

Difference and Transformer-based models achieve the best results. 19

Keywords: Ultrasound imaging ; Image Segmentation ; Intima-media thickness ; Carotid artery ; 20

Deep Learning 21

0. Introduction 22

The primary mechanism in the human body that sustains life is the cardiovascular 23

system. Cardio-vascular system diseases (CVDs) have been regarded as a major cause 24

of death in the world. Lifespan can be increased, and the death rate from CVDs can be 25

decreased with early diagnosis and treatment of the diseases. The cardiovascular system is 26

made up of blood vessels that carry blood, which is necessary for all of the body’s organs 27

to operate. The primary components of the blood vessels that transport blood to and from 28

the heart to all organs are arteries and veins. Any obstruction of blood flow or disease in 29

the arteries or veins seriously affects how well the organs it supports operate. The most 30

common types of cardiovascular disease include peripheral vascular disease, coronary 31

artery disease, and Carotid artery disease. These disorders manifest as a result of the 32

development of atherosclerotic plaque in the arteries as illustrated in Figure 1. One of 33

the effects of carotid artery stenosis, the accumulation of plaque on the carotid artery, is 34

ischemic stroke. If the stenosis is detected early and the amount of plaque is determined, 35

the problem can be addressed immediately. For this, a variety of imaging modalities are 36

used. Computed tomography (CT), EEG, ECG, ultrasound imaging, laboratory tests for 37

coagulation status, and cardiac monitoring are among the diagnostic techniques used in 38

the assessment of carotid artery stenosis or the assessment of stroke. Both sides of the 39

neck contain the common carotid artery. The soft tissue features in the arteries allow for 40

imaging using a variety of methods, or modalities, such as computerized tomography (CT), 41

Ultrasound imaging, and magnetic resonance imaging. The analysis of generated images 42

can enhance diagnosis and support clinical judgment. Medical image analysis algorithms 43

have advanced significantly, from image processing and pattern recognition methods to 44

machine learning and deep learning algorithms that see it as a computer vision problem. 45

A notable development in automatic segmentation, analysis, and grading of stenosis has 46

been witnessed with the use of carotid artery imaging generated by CT scans, MRIs, and 47

Appl. Sci. 2022, 1, 0. https://doi.org/10.3390/app1010000 https://www.mdpi.com/journal/applsci

Appl. Sci. 2022, 1, 0 2 of 11

Figure 1. Visualization of plaque builds up and blocks the normal flow of blood in the artery (https: //my.clevelandclinic.org/health/diseases/16845-carotid-artery-disease-carotid-artery-stenosis, ac- cessed on 9 January 2023).

Ultrasound images [1], [2]. Due to the complexity of scanning the carotid artery ultrasound 48

scanning is the preferred way to capture images with acceptable resolutions. Ultrasound 49

images have been used for many studies using medical imaging analysis algorithms. 50

In order to segment the plaques on the carotid artery, many methods have been 51

proposed even with the lack of large-scale datasets. Before, the proposed methods used 52

cIMT measurement to detect and localize the region of carotid artery walls and then 53

the plaques [3,4]. The ground truth was presented using some points that represent the 54

generated by specialists [5]. The analysis using these types of data used different statistical 55

and machine learning algorithms including snake’s segmentation and contour [3,4], bulb 56

edge detection [5], wind-driven optimization techniques [6], and SVM [7]. 57

Using convolution neural networks, the proposed methods used binary segmentation 58

instead of cIMT measurement. By generating binary images that contain labeled regions 59

in the images instead of using points, the deep learning methods succeed to segment 60

these regions with better precision [7]. Also, the segmented regions can be helpful to 61

compute Carotid Intima-Media Thickness [8] that is related to the performance accuracy of 62

segmentation. That makes the segmentation part a crucial task. 63

Although CNNs have succeeded in solving many computer vision problems, recent 64

studies have shown many drawbacks for CNNs such as the greedy need for large-scale 65

datasets [9], and the reliance on linear neuron models [10–13]. Operational Neural Net- 66

works (ONNs) [13–16] are heterogeneous networks with a non-linear neuron model that 67

are recently proposed as a solution for highly non-linearly separable problems. With the 68

help of predefined nodal, pool, and activation operators, ONNs are able to learn highly 69

complex and multi-modal functions. The transformer neural network has recently been 70

a successful non-convolutional neural network alternative for computer vision problems. 71

Instead of convolution, vision transformers utilize self-attention to combine information 72

from several locations [41]. In this paper, we performed a segmentation of Common 73

Carotid Intima-Media using deep learning models. For that, we updated some existing 74

deep learning models like DeepCrack [42] and the Transformer-based model [43]. We used 75

Self-ONN instead of normal convolutional layers for DeepCrack. In order to improve the 76

segmentation quality we used some morphological operations like erosion to enhance the 77

output results. The main contributions of the research are summarized as the following: 78

• We develop and investigate various recent deep-learning models for the segmentation 79

of IMC in the B-mode ultrasound images of the carotid artery 80

Appl. Sci. 2022, 1, 0 3 of 11

• We propose a pioneer application for self-organized operational neural networks 81

(Self-ONNs) for IMC segmentation 82

• We investigate the level of non-linearity for operational layers required to achieve 83

better segmentation performance. 84

The rest of the paper is divided as follows; in Section 2, we highlight the recent work of 85

carotid intima-media segmentation. Then, in Section 3, we present the model architecture 86

for the deep learning models. Whereas, in Section 4, we present the experimental setup 87

along with the evaluation metrics and the results of the model. Finally, we conclude and 88

explain the future work in Section 5. 89

1. Related works 90

The carotid artery segmentation including walls and the plaques in the intima-media 91

complex (IMC) can be used for the estimation of intima-media thickness (IMT). Which 92

makes it an important operation of risk evaluation for atherosclerosis development. 93

There are numerous methods for segmenting the intima-media complex. However, 94

the majority of them are semi-automatic and require manual intervention. Medical experts 95

define the boundary between the media adventitia and lumen. However, the subjectivity 96

and variability of manual segmentation can be reduced using image segmentation algo- 97

rithms. Additionally, Intima-Media Thickness (IMT) is assessed using active contours 98

[17–24], dynamic programming [25–30], and edge detection algorithms and gradient-based 99

approaches [31,32]. For active contours-based approaches, the authors in [17] start with a 100

simple segmentation of the B-mode ultrasound images followed by Segmentation of far 101

wall intima-media-adventitia the applied the active contour to get the desired region in the 102

images. The same process in [18] but this time using some morphological operation like 103

opening then a LI contour function is applied to detect the final Common Carotid Artery 104

result. In [19], the authors started with non-linear filtering followed by the detection of 105

the intima layer using an iterative relaxation procedure and then detecting the wall using 106

modified energy functional and an optimal initial contour. 107

For dynamic programming-based approaches, the researchers in [25] used multi-scale 108

dynamic programming (DP) algorithm to estimate the vessel wall positions that lead to 109

the detection of the boundaries. the obtained results with geometrical characteristics are 110

used to get the final results. in the same context, and for Detecting the Arterial Wall, the 111

authors in [27] proposed a dual dynamic programming (DDP) technique for detecting 112

intima and adventitial layers of the common carotid artery. Also in [29] an improved 113

dynamic programming method is proposed for carotid artery wall thickness evaluation. 114

With the introduction of the machine and deep learning techniques become interested 115

as a promising method for medical image analysis tasks such as image denoising, segmenta- 116

tion, and classification. Before the development of deep learning models, machine learning 117

was the most commonly utilized technology, where comprehensive feature extraction 118

techniques were applied to find several areas of carotid artery risk estimation. The deep 119

learning strategy takes advantage of a neural network architecture that mimics the human 120

brain by having more hidden layers. The neuron is the fundamental building block of a 121

deep neural network (DNN), which accepts several inputs, linearly combines them, and 122

then passes them to a nonlinear network to produce the desired output. Multiple process- 123

ing layers make up a deep learning network, which uses deep graphs to extract high-level 124

representations of meaningful information from low-level inputs. Convolutional Neural 125

Networks are among the widely used networks in the medical image analysis domain 126

[33]. The U-Net is a CNN-based architecture used to solve the problem of automatic image 127

segmentation . This architecture is adopted in many IMC segmentation research such as 128

[34–36]. For example, in [34] the authors used UNet architecture for plaque segmentation 129

in carotid Ultrasound images. In [37], UNet is also used for the same purpose. Also, in [38] 130

the authors used UNet, UNet+, UNet++, and UNet+++, and three types of hybrids, namely, 131

Inception-UNet, Fractal-UNet, and Squeeze-UNet architectures to segments and measures 132

the area of the plaque far wall of the common carotid (CCA) and internal carotid arteries 133

Appl. Sci. 2022, 1, 0 4 of 11

(a) Self-ONN-DeepCrack

(b) Transformer Network

Figure 2. The networks used for Ultrasound IMC segmentation

(ICA) in B- mode ultrasound. Using M-Net [39] as backbone, the authors in [40] proposed 134

Automatic joint segmentation method named CSM-Net with Triple spatial attention and 135

cascaded dilated convolution modules. 136

2. Proposed Approach 137

The proposed methodology uses deep learning methods for Ultrasound IMC seg- 138

mentation. We used these methods for comparison as well as we adapted some of them 139

to be suitable for our purpose. The used deep learning methods including DeepCrack 140

[43], PidiNet [9], and CCTrans [43] are used for different tasks like edge detection, Crack 141

Segmentation, and crowd counting. For the DeepCrack network, we adapted the model 142

with Self-ONNs techniques. CCTrans is a transformer-based model used for crowd count- 143

ing. For that, we adapted the model to be suitable for Ultrasound IMC segmentation by 144

exploiting the same first layers of the model. 145

Appl. Sci. 2022, 1, 0 5 of 11

2.1. Self-Operational Neural Network-based Model 146

Self-Organized Operational Neural Networks with Generative Neurons, proposed by [46], are a type of artificial neural network that is designed to operate in a self-organizing manner. Instead of using a predefined set of operators as ONNs, the Self-ONNs with generative neurons generate nodal operators during back-propagation training. This property of Self-ONNs allows for maximum learning performance, diversity, and flexibility. The use of generative neurons can improve the network’s robustness to unseen data and reduce the risk of over fitting. A generative neuron uses a Taylor series expansion around the point a to approximate the nonlinear function f(x):

Y = S

∑ s=1

f n(a) n!

(x − a)2 (1)

If we truncate the Taylor series to q terms then the approximation g(w, x, a) will be 147

given by: 148

Y = w0 + w1(x − a) + ... + wq(x − a)q (2)

Where wn = f n(a) n! (x − a)2, The w0 is the bias and for the c-channel input tensor, the 149

wn, n = 1,. . . ,q is the q banks of c-channel convolution kernels that are learned during 150

back-propagation. 151

To investigate the performance of Self-ONNs we choose the DeepCrack [42] model as 152

a baseline model. The DeepCrack network proposed by [42] is a CNN-based model built 153

for crack segmentation. The architecture of the DeepCrack network is shown in Figure 154

1. It has thirteen convolutional layers, each has convolution, batch normalization, and 155

ReLU layers. The convolution produces a set of feature maps. At the same time, batch 156

normalization is used to reduce the covariate shift, and the ReLU is the activation function 157

used to learn non-linearity in the data. A Max-pooling with 2 × 2 pixel filter layers is 158

added between the convolutional layers. A convolutional layer with kernel size 1 is used 159

to obtain side-output features. Deconvolutional layers are then used (Except for the layer 160

of the first side output ) to upsample the feature maps’ plane size to match the input image. 161

Following the concatenation of the up-sampled feature maps to get the final features, a 162

convolutional layer and a softmax layer are applied. Then a convolutional layer followed 163

by a softmax layer for predicting 2 classes is used. According to this prediction, for each 164

pixel, the predicted label can be obtained. We modified the network to be flexible to use 165

Self-ONN layers instead of CNN layers, as shown in Figure 3. We use Tanh activation layers 166

instead of ReLU. The level of non-linearity can be adjusted on the network by modifying 167

the parameter q. 168

2.2. Pixel Difference-based Model 169

Although CNNs can achieve human-level performance in many computer-Vision- 170

based applications, the high performance of CNN-based models is achieved with a large 171

pre-trained CNN backbone like VGG, ResNet, and DenseNet, which is memory and energy- 172

consuming. While some methods proposed simple and lightweight architectures like Pixel 173

Difference Network (PiDiNet) that use edge detection [9]. PiDiNet adopts novel pixel 174

difference convolutions that integrate the traditional edge detection operators into the 175

popular convolutional operations in modern convolution neural networks for enhanced 176

performance on the task, which enjoys the best of both worlds. We used PidiNet model 177

other for IMC segmentation. 178

2.3. Transformer-based Model 179

CNNs have a strong ability to extract local features, but they inherently fail in modeling 180

the global context due to the limited receptive fields. The transformer can model the global 181

context easily. Also, it has become the most used technique in computer vision. For 182

Appl. Sci. 2022, 1, 0 6 of 11

(a) Ground-truth (b) Segmentation result (b) Erosion result

Figure 3. The morphological erosion on Ultrasound IMC Segmentation results

that, we used a transformer model for IMC segmentation. he proposed method used a 183

pyramid vision transformer backbone to capture the global information, a pyramid feature 184

aggregation (PFA) model to combine low-level and high-level features, and an efficient 185

regression head with multi-scale dilated convolution (MDC) to predict the final results 186

[43]. The input image is transformed into a 1D sequence first, then the output is fed into 187

the transformer-based backbone. The Pyramid transformer in [44] is adopted to capture 188

the global context through various downsampling stages. The outputs of each stage are 189

reshaped into 2D feature maps for pyramid feature aggregation. Finally, a simple regression 190

head with multi-scale receptive fields regresses the final results. The proposed architecture 191

is illustrated in Figure 3. 192

2.4. Post-Processing 193

The Carotid Intima-Media Thickness is a difficult task using the image segmentation 194

method, due to the difficulty of generating the precise thickness from an image using even 195

deep learning methods. While the Carotid Intima-Media region can be segmented but for 196

some images, this region can be very skinny which can affect the performance of such a 197

method. Also using deep learning methods the segmented Thickness is generally fat as 198

presented in Figureerosion (b). For that, and in order to make the segmented thickness 199

skinny to meet the ground truth we applied morphological erosion. An example of the 200

erosion result is presented in Figureerosion (c). 201

3. Experimental results 202

In this section, we demonstrate the experimental results of the proposed Slef-ONN- 203

DeepCrack approach on the CUBs dataset, as well as compared the obtained results with 204

some image segmentation methods in the literature including DeepCrack [43], PidiNet [9], 205

and adapted CCTrans [44]. The comparison has been performed using image segmentation 206

metrics as well as visual illustration. 207

3.1. Dataset and evaluation metrics 208

The dataset used in this study is the CUBs dataset which is acquired from both sides of the neck of 1088 participants. The total is 2176 images. All images are annotated by a skilled analyst. The images in Figure 5. are samples of the images and the ground truths taken from the dataset. The segmentation metrics used to evaluate the performance of the proposed models are Precision, Recall, F1-Measure (Equation 3), Jaccard Index (Equation 4), and Dice coefficient (Equation 5). The Precision measures how many True Positive (TP) predictions there are out of all the positive ones, or how many positive predictions there are in total. The Recall is the True Positive Rate (TPR), or how many predictions of true

Appl. Sci. 2022, 1, 0 7 of 11

Figure 4. The Precision-Recall curve of Ultrasound IMC segmentation.

Table 1. Performance of the proposed and the implemented models On CUBS dataset

Model Precision Recall F-Measure Dice Jaccard DeepCrack_CNN 0.631 0.675 0.652 0.652 0.484 DeepCrack_CNN + post-proc 0.834 0.618 0.697 0.697 0.544 DeepCrack_Self (Q=3) 0.652 0.688 0.669 0.669 0.503 DeepCrack_Self + post-proc 0.792 0.691 0.721 0.721 0.571 PiDiNet 0.687 0.825 0.750 0.750 0.60 PiDiNet + post-proc 0.876 0.740 0.791 0.791 0.661 Transformer 0.68 0.826 0.746 0.746 0.595 Transformer + post-proc 0.882 0.849 0.801 0.801 0.656

positives are made out of all the Actual Positives. Both precision and recall are used to handle the class imbalance problem and they are used to compute the F1 score.

F1 = 2 ∗ Precision ∗ Recall

Precision + Recall (3)

Jaccard Index = True Positive

True Positive + False Negative + False Positive (4)

Dice = 2 ∗ 2True Positive

2 ∗ True Positive + False Negative + False Positive (5)

3.2. Evaluation 209

In order to evaluate the Ultrasound IMC segmentation using the deep learning meth- 210

ods on CUBs dataset, a set of metrics are used including model Precision, Recall, F-measure, 211

Dice, and Jaccard. These metrics are generally the most used ones for image segmentation 212

in computer vision tasks. In this section, we present the obtained results per dataset using 213

the cited method for Ultrasound IMC segmentation. These results are reported in Tables 214

and figures in order to show the most performed techniques using different architectures. 215

We first investigate the effect of replacing CNN layers with Self-ONN layers in the 216

DeepCrack model. The level of linearity is controlled using the parameter q=3,5,7,9,11. Fig- 217

ure 4 (Left) shows that the best-performing model uses q=3 then the accuracy of the model 218

starts to drop as we enlarge the level of non-linearity. Compared with the CNN version 219

of the model, Figure 4 (Middle) shows that the best precision and recall accuracies were 220

better when q is set to 3 and 5. The performance of all deep learning models on the CUBS 221

dataset is shown in Table 1. From the table, we can observe that both Transformer-based 222

and Pixel Difference-based models act similarly in terms of all performance measures with 223

a slight increase for PiDiNet in the F-measure, Dice, and Jaccard index. Both Trans-former- 224

Based and Pixel Difference-Based achieved better performance with exceptional margins 225

compared to CNN and Self-ONN-based models. From Table 1 We can find also that the 226

post-processing operations improve the performance metrics of all methods including Deep- 227

Appl. Sci. 2022, 1, 0 8 of 11

Figure 5. Original and ground truth sample images and the corresponding segmentation results for the proposed deep learning models.

Crack, DeepCrack_Self_ONN, PidiNet, and Transformer-based models. The improvement 228

achieved about 20%, 14%, 19%, and 20% for DeepCrack, DeepCrack_Self_ONN, PidiNet, 229

and Transformer-based models respectively on Precision metrics. While the Transformer- 230

based + post-pro model reached the best metrics values followed by PidiNet + post-pro 231

with an average difference of 1%, 10%, and less than 1% for Dice, Recall, and Precision 232

respectively. in addition to qualitative results we present the qualitative results presented 233

in Figure 5 that, show some visual output for the segmentation results. For Figure 5 we can 234

see that all the proposed methods have reached a segmentation with a good performance, 235

with a sample difference in terms of thickness. 236

4. Conclusion 237

We developed and investigated various recent deep-learning models for the segmen- 238

tation of IMC in the B-mode ultrasound images of the carotid artery. Compared to the 239

conventional CNN-based model, the Self-ONN-Based model performs better in all evalu- 240

ation metrics, however, the Pixel Difference and the Transformer-based models perform 241

better in all metrics and that might be due to the lack of enough data. The Pixel Difference 242

model performs better in case the data is scarce. A further investigation of suitable data 243

augmentation techniques is needed to raise the accuracy. 244

Appl. Sci. 2022, 1, 0 9 of 11

Author Contributions: Conceptualization, A.S., O.E.,and S.A.-M.; data curation, A.S.; formal analysis, 245

A.S.; methodology, A.S., O.E., and S.A.-M.; project administration, S.A, and N.A.; supervision, S.A.- 246

M., and N.A.; validation, A.S., O.E., S.A.-M., and N.A.; visualization, A.S., and O.E.; writing—original 247

draft, A.S.; writing—review and editing, A.S., O.E., S.A.-M., and N.A. All authors have read and 248

agreed to the published version of the manuscript.. 249

Institutional Review Board Statement: Not applicable. “Not applicable” for studies not involving 250

humans or animals. 251

Informed Consent Statement: Not applicable 252

Data Availability Statement: Not applicable 253

Acknowledgments: This publication was supported by Qatar University Internal Grant QUHI- 254

CENG-22/23-548, respectively. The findings achieved herein are solely the responsibility of the 255

authors. 256

Conflicts of Interest: The authors declare no conflict of interest. 257

Sample Availability: Samples of the compounds ... are available from the authors. 258

Abbreviations 259

The following abbreviations are used in this manuscript: 260

261

DR Diabetic Retinopathy DL Deep learning AI Artificial Intelligence CNN Convolutional Neural Network

262

References 263

1. S. Latha, D. Samiappan, and R. Kumar, “Carotid artery ultrasound image analysis: A review of the literature,” Proc Inst Mech Eng 264

H, vol. 234, no. 5, pp. 417–443, 2020. 265

2. M. del M. Vila, B. Remeseiro, M. Grau, R. Elosua, and L. Igual, “Last Advances on Automatic Carotid Artery Analysis in Ultrasound 266

Images: Towards Deep Learning,” in Handbook of Artificial Intelligence in Healthcare, Springer, 2022, pp. 215–247. 267

3. Loizou, C.; Kasparis, T.; Spyrou, C.; Pantzaris, M. Integrated system for the complete segmentation of the common carotid artery 268

bifurcation in ultrasound images. In IFIP International Conference on Artificial Intelligence Applications and Innovations; Springer: 269

Berlin/Heidelberg, Germany, 2013; Volume 412, pp. 292–301. 270

4. Christodoulou, L.; Loizou, C.P.; Spyrou, C.; Kasparis, T.; Pantziaris, M. Full-automated system for the segmentation of the common 271

carotid artery in ultrasound images. In Proceedings of the 2012 5th International Symposium on Communications, Control and 272

Signal Processing, Rome, Italy, 2–4 May 2012; pp. 1–6. 273

5. Ikeda, N.; Dey, N.; Sharma, A.; Gupta, A.; Bose, S.; Acharjee, S.; Shafique, S.; Cuadrado-Godia, E.; Araki, T.; Saba, L.; et al. 274

Automated Segmental-IMT Measurement in Thin/Thick Plaque with Bulb Presence in Carotid Ultrasound from Multiple Scanners: 275

Stroke Risk Assessment. Comput. Methods Programs Biomed. 2017, 141. 276

6. Madipalli, P.; Kotta, S.; Dadi, H.; Nagaraj, Y.; Asha, C.S.; Narasimhadhan, A.V. Automatic Segmentation of Intima Media Complex 277

in Common Carotid Artery using Adaptive Wind Driven Optimization. In Proceedings of the 2018 Twenty Fourth National 278

Conference on Communications (NCC), Hyderbad, India, 25–28 February 2018; pp. 1–6. 279

7. Nagaraj, Y.; Teja, A.; Narasimha, D. Automatic Segmentation of Intima Media Complex in Carotid Ultrasound Images Using 280

Support Vector Machine. Arab. J. Sci. Eng. 2018, 44. 281

8. Biswas, M., Saba, L., Chakrabartty, S., Khanna, N. N., Song, H., Suri, H. S., ... Suri, J. S. (2020). Two-stage artificial intelligence 282

model for jointly measurement of atherosclerotic wall thickness and plaque burden in carotid ultrasound: A screening tool for 283

cardiovascular/stroke risk assessment. Computers in biology and medicine, 123, 103847. 284

9. Z. Su et al., “Pixel difference networks for efficient edge detection,” in Proceedings of the IEEE/CVF International Conference on 285

Computer Vision, 2021, pp. 5117–5127. 286

10. S. Kiranyaz, J. Malik, H. ben Abdallah, T. Ince, A. Iosifidis, and M. Gabbouj, “Self-organized Operational Neural Networks with 287

Generative Neurons,” Neural Networks, vol. 140, pp. 294–308, Aug. 2021, doi: 10.1016/j.neunet.2021.02.028. 288

11. M. Gabbouj et al., “Robust Peak Detection for Holter ECGs by Self-Organized Operational Neural Networks,” IEEE Trans Neural 289

Netw Learn Syst, pp. 1–12, 2022, doi: 10.1109/TNNLS.2022.3158867. 290

12. J. Malik, S. Kiranyaz, and M. Gabbouj, “Operational vs Convolutional Neural Networks for Image Denoising,” arXiv preprint 291

arXiv:2009.00612, 2020. 292

13. J. Malik, O. C. Devecioglu, S. Kiranyaz, T. Ince, and M. Gabbouj, “Real-Time Patient-Specific ECG Classification by 1D Self- 293

Operational Neural Networks,” IEEE Trans Biomed Eng, pp. 1–1, 2021, doi: 10.1109/TBME.2021.3135622. 294

Appl. Sci. 2022, 1, 0 10 of 11

14. A. Rahman et al., “Robust biometric system using session invariant multimodal EEG and keystroke dynamics by the ensemble of 295

self-ONNs,” Computers in Biology and Medicine, vol. 142, p. 105238, Mar. 2022, doi: 10.1016/j.compbiomed.2022.105238. 296

15. M. Soltanian, J. Malik, J. Raitoharju, A. Iosifidis, S. Kiranyaz, and M. Gabbouj, “Speech Command Recognition in Computationally 297

Constrained Environments with a Quadratic Self-Organized Operational Layer,” in 2021 International Joint Conference on Neural 298

Networks (IJCNN), Jul. 2021, pp. 1–6. doi: 10.1109/IJCNN52387.2021.9534232. 299

16. S. Kiranyaz, T. Ince, A. Iosifidis, and M. Gabbouj, “Operational neural networks,” Neural Comput Appl, vol. 32, no. 11, pp. 300

6645–6668, Jun. 2020, doi: 10.1007/s00521-020-04780-3. 301

17. S. Petroudi, C. Loizou, M. Pantziaris, and C. Pattichis, “Segmentation of the common carotid intima-media complex in ultrasound 302

images using active contours,” IEEE Trans Biomed Eng, vol. 59, no. 11, pp. 3060–3069, 2012. 303

18. F. Molinari et al., “Constrained snake vs. conventional snake for carotid ultrasound automated IMT measurements on multi-center 304

data sets,” Ultrasonics, vol. 52, no. 7, pp. 949–961, 2012. 305

19. M. Ceccarelli, N. De Luca, and A. Morganella, “An active contour approach to automatic detection of the intima-media thickness,” 306

in 2006 IEEE International Conference on Acoustics Speech and Signal Processing Proceedings, 2006, vol. 2, pp. II–II. 307

20. C. P. Loizou, C. S. Pattichis, M. Pantziaris, T. Tyllis, and A. Nicolaides, “Snakes based segmentation of the common carotid artery 308

intima media,” Med Biol Eng Comput, vol. 45, no. 1, pp. 35–49, 2007. 309

21. M. A. Gutierrez, P. E. Pilon, S. G. Lage, L. Kopel, R. T. Carvalho, and S. S. Furuie, “Automatic measurement of carotid diameter and 310

wall thickness in ultrasound images,” in Computers in Cardiology, 2002, pp. 359–362. 311

22. R. C. Chan, J. Kaufhold, L. C. Hemphill, R. S. Lees, and W. C. Karl, “Anisotropic edge-preserving smoothing in carotid B-mode 312

ultrasound for improved segmentation and intima-media thickness (IMT) measurement,” in Computers in Cardiology 2000. Vol. 27 313

(Cat. 00CH37163), 2000, pp. 37–40. 314

23. S. Delsanto, F. Molinari, P. Giustetto, W. Liboni, S. Badalamenti, and J. S. Suri, “Characterization of a completely user-independent 315

algorithm for carotid artery segmentation in 2-D ultrasound images,” IEEE Trans Instrum Meas, vol. 56, no. 4, pp. 1265–1274, 2007. 316

24. J. H. Gagan et al., “Automated Segmentation of Common Carotid Artery in Ultrasound Images,” IEEE Access, 2022. 317

25. Q. Liang, I. Wendelhag, J. Wikstrand, and T. Gustavsson, “A multiscale dynamic programming procedure for boundary detection 318

in ultrasonic artery images,” IEEE Trans Med Imaging, vol. 19, no. 2, pp. 127–142, 2000. 319

26. I. Wendelhag, Q. Liang, T. Gustavsson, and J. Wikstrand, “A new automated computerized analyzing system simplifies readings 320

and reduces the variability in ultrasound measurement of intima-media thickness,” Stroke, vol. 28, no. 11, pp. 2195–2200, 1997. 321

27. D.-C. Cheng and X. Jiang, “Detections of arterial wall in sonographic artery images using dual dynamic programming,” IEEE 322

Transactions on Information Technology in Biomedicine, vol. 12, no. 6, pp. 792–799, 2008. 323

28. T. Gustavsson, Q. Liang, I. Wendelhag, and J. Wikstrand, “A dynamic programming procedure for automated ultrasonic measure- 324

ment of the carotid artery,” in Computers in Cardiology 1994, 1994, pp. 297–300. 325

29. N. Santhiyakumari and M. Madheswaran, “Non-invasive evaluation of carotid artery wall thickness using improved dynamic 326

programming technique,” Signal Image Video Process, vol. 2, no. 2, pp. 183–193, 2008. 327

30. Y.-B. Lee, Y.-J. Choi, and M.-H. Kim, “Boundary detection in carotid ultrasound images using dynamic programming and a 328

directional Haar-like filter,” Comput Biol Med, vol. 40, no. 8, pp. 687–697, 2010. 329

31. C. Liguori, A. Paolillo, and A. Pietrosanto, “An automatic measurement system for the evaluation of carotid intima-media 330

thickness,” IEEE Trans Instrum Meas, vol. 50, no. 6, pp. 1684–1691, 2001. 331

32. R. H. Selzer, W. J. Mack, P. L. Lee, H. Kwong-Fu, and H. N. Hodis, “Improved common carotid elasticity and intima-media 332

thickness measurements from computer analysis of sequential ultrasound frames,” Atherosclerosis, vol. 154, no. 1, pp. 185–193, 333

2001. 334

33. A. S. Pramulen, E. M. Yuniarno, J. Nugroho, I. M. G. Sunarya, and I. K. E. Purnama, “Carotid Artery Segmentation on Ultrasound 335

Image using Deep Learning based on Non-Local Means-based Speckle Filtering,” in 2020 International Conference on Computer 336

Engineering, Network, and Intelligent Multimedia (CENIM), 2020, pp. 360–365. 337

34. P. K. Jain et al., “Automated deep learning-based paradigm for high-risk plaque detection in B-mode common carotid ultrasound 338

scans: An asymptomatic Japanese cohort study,” Int. Angiol, vol. 41, pp. 9–23, 2021. 339

35. N. Lainé, G. Zahnd, M. Orkisz, and others, “Carotid artery wall segmentation in ultrasound image sequences using a deep 340

convolutional neural network,” arXiv preprint arXiv:2201.12152, 2022. 341

36. N. Radovanović, L. Dašić, A. Blagojević, T. Šušteršič, and N. Filipović, “Carotid Artery Segmentation Using Convolutional Neural 342

Network in Ultrasound Images”. 343

37. Park, J. H., Seo, E., Choi, W., Lee, S. J. (2022). Ultrasound deep learning for monitoring of flow–vessel dynamics in murine carotid 344

artery. Ultrasonics, 120, 106636. 345

38. Jain, P. K., Sharma, N., Kalra, M. K., Johri, A., Saba, L., Suri, J. S. (2022). Far wall plaque segmentation and area measurement in 346

common and internal carotid artery ultrasound using U-series architectures: An unseen Artificial Intelligence paradigm for stroke 347

risk assessment. Computers in Biology and Medicine, 149, 106017. 348

39. H. Fu, J. Cheng, Y. Xu, D.W.K. Wong, J. Liu, X. Cao, Joint optic disc and cup segmentation based on multi-label deep network and 349

polar transformation, IEEE Trans. Med. Imag. 37 (7) (2018) 1597–1605 350

40. Yuan, Y., Li, C., Xu, L., Zhu, S., Hua, Y., Zhang, J. (2022). CSM-Net: Automatic joint segmentation of intima-media complex and 351

lumen in carotid artery ultrasound images. Computers in Biology and Medicine, 150, 106119. 352

Appl. Sci. 2022, 1, 0 11 of 11

41. S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, “Transformers in vision: A survey,” ACM computing surveys 353

(CSUR), vol. 54, no. 10s, pp. 1–41, 2022. 354

42. Y. Liu, J. Yao, X. Lu, R. Xie, and L. Li, “DeepCrack: A deep hierarchical feature learning architecture for crack segmentation,” 355

Neurocomputing, vol. 338, pp. 139–153, 2019. 356

43. Liu, Y., Yao, J., Lu, X., Xie, R., Li, L. (2019). DeepCrack: A deep hierarchical feature learning architecture for crack segmentation. 357

Neurocomputing, 338, 139-153. 358

44. X. Chu et al., “Twins: Revisiting the design of spatial attention in vision transformers,” Adv Neural Inf Process Syst, vol. 34, pp. 359

9355–9366, 2021. 360

45. K. M. Meiburger et al., “Carotid ultrasound boundary study (CUBS): an open multicenter analysis of computerized intima–media 361

thickness measurement systems and their clinical impact,” Ultrasound Med Biol, vol. 47, no. 8, pp. 2442–2455, 2021. 362

46. Kiranyaz, S., Malik, J., Abdallah, H. B., Ince, T., Iosifidis, A., Gabbouj, M. (2021). Self-organized operational neural networks with 363

generative neurons. Neural Networks, 140, 294-308. 364