Automatic Diagnosis of Common Carotid Artery and Plaque Grading Using Novel Deep Learning Model

Several methods have been developed for CCA segmentation and diagnosis using different strategies. This section provides a review of recent works related to CCA segmentation and classification.

For tasks involving medical image analysis, including preprocessing, segmentation, and diagnosis, machine learning (ML) and deep learning (DL) algorithms have emerged as intriguing and promising approaches. DL based Intima-media Complex (IMC) segmentation was described by Mohammed et al. [1]. Four DL models, such as Convolutional Neural Network (CNN), transformer-based, pixel difference-based, and self-organizing operational neural networks, were employed for isolating IMC from the CUBS data. Results demonstrated that deep learning models can offer significant improvements in IMC segmentation. However, deep learning models require extensive data to achieve satisfactory results.

Menchón-Lara and Sancho-Gómez [3] presented an ML model for IMT division in American images. From his perspective, the US images were prepared to identify ROI. To each pixel, to produce an intensity pattern, a windowing approach to ROI was used. These intensity patterns were represented as a feature space using an autoencoder. An Artificial Neural Network (ANN) was used to identify Lumen Intima (Li) and Media Adventitia (MA) interfaces. However, this method's dependence on an autoencoder can cause information loss during a decrease in diarrhea, potentially affecting the division's accuracy.

Nagaraj et al. [7] used a Wind Driven Optimization (WDO) method to fragment IMT depending on a certain point. To remove noise, images were denoised using an Optimized Bayesian Non-Local Mean (OBNLM) filter. After this, total variation followed by morphological operations was employed to improve visual appearance and clean the borders and edges. A Canny edge detector was applied to detect edges. Finally, a threshold-based WDO algorithm was implemented to estimate IMT. This method may not generalize well across different US systems.

Rocha et al. [14] showed a way to divide US images into near-end and far-end intima regions. The authors were able to make accurate predictions for IMT by using geometric snakes, dynamic programming, and smooth intensity thresholding. Geometric snakes may use a significant amount of processing power and be sensitive to where the beginning contour is positioned. This might cause problems when the intima layers are not perfectly constructed.

Naik et al. [15] provided a thorough analysis of methods available for carotid artery segmentation in US images, along with their merits and weaknesses. A data mining model for plaque grading was introduced by Molinari et al. [16]. US images were preprocessed using CLATHE and then decomposed using the Bidirectional Empirical mode decomposition method. Energy features were computed, and then the dimension of features was reduced using some tests, such as the t-test and the Wilcoxon test. Finally, Linear Discriminant Analysis (LDA) was adopted to differentiate between plaque that is asymptomatic and that is symptomatic. This method may suffer from computational inefficiency due to feature decomposition and preprocessing steps.

Latha et al. [17] recommended an automatic segmentation method for carotid Artery segmentation. US images were denoised with the curvelet transform with the soft thresholding technique. Affinity Propagation followed by Density-Based Spatial Clustering of Applications with Noise (AB+DBSCAN) was adopted for segmentation. Results were compared with the PSO and IDO algorithms. Clustering methods like AB+DBSCAN may struggle with noise and artifacts, leading to inconsistent segmentation results.

Vila et al. [18] suggested a method based on DenseNet for plaque segmentation and IMT estimation. This method does not need any ROI extraction. Input images were applied to a densely connected CNN for semantic segmentation. After segmentation, IMT is estimated, and plaque is detected. Deep learning-based models are resource-intensive, requiring substantial computational power. Huang et al. [19] provided a comprehensive survey of CCA disease diagnosis methods using DL, along with their merits and drawbacks.

He et al. [20] designed a Bilinear CNN with a residual Neural network (BCNN-Resnet) to recognize carotid plaques. To eliminate unnecessary background, input images were cropped to a uniform size. Flipping, scaling, and rotation were some of the augmentation methods employed to expand the database's ability. BCNN-ResNet was applied to perform classification task. Resizing images could potentially distort anatomical structures, impacting segmentation accuracy.

Kumar et al. [21] used active oblongs to segment CCA images. But an active oblongs-based segmentation method may struggle with anatomical variations and irregularities. Jain et al. [22] developed an automated system using DL and reported satisfactory results. Özdemir et al. [23] explored an ML model for carotid artery disease diagnosis. High classification accuracy was attained by the technique using a super learning model that combines random forest, gradient boosting, and adaptive boosting. Although this method increased the size of the dataset to 1000 by means of an augmentation method. The original images were 122, relatively small. This may limit the generalizability of the model.

CCA disease is a serious global medical problem, and prevention depends on early disease recognition. Fitas and Goncalves [24] investigated how to enhance data for the detection of carotid artery disease using Generative Adversarial Networks (GAN). Both original and enhanced data were used to confirm the classification rate of four deep learning models: AlexNet, VGGNet, GoogleNET, and CifarNet. Results highlighted the power of GAN in improving deep learning models for medical image analysis. Nevertheless, training a GAN is computationally intensive and can be challenging to stabilize, requiring significant computational resources.

Koloi et al. [25] predicted early-stage coronary artery disease using a machine learning model. The authors considered 331 original images and then applied some techniques to generate high-quality virtual data to improve the machine learning model performance. Gradient boosting showed better performance than random forest in terms of accuracy. Ottakath et al. [26] presented a Bi-attention doubleUNet for carotid artery segmentation. This approach combined channel-wise and spatial attention with a bottleneck module. The model was tested on US images and reported superior segmentation performance. The drawback of the method is that the effectiveness relies on high-quality US images.

2.1 Research issue and contributions

This paper presents significant advancements in the development of the CADS for carotid artery disease. The system is distinct in implementing a five-phase process, which sets it apart from previous methods in this domain. The primary contributions of this work are outlined across several key dimensions:

First, this research presents a novel CADS system that incorporates a series of sophisticated processes not previously combined in this manner. Each phase contributes to the overall innovation:

Preprocessing: A new CMF is proposed to effectively eliminate noise while preserving diagnostic information, unlike earlier methods that relied on standard filters.

Segmentation: ROI extraction is performed using a threshold-based metaheuristic algorithm, Modified AOA, offering improved precision compared to conventional segmentation techniques.

Disease detection: A ViT is employed as the first-level classifier, diverging from the typical use of CNN or traditional ML models.

Plaque grading: A SVM serves as the second-level classifier, enabling quantification of plaque severity. To the best of the authors’ knowledge, this is the first attempt to apply a two-level classification strategy for CCA diagnosis.

Another key contribution lies in integrating preprocessing, segmentation, and a two-level classification pipeline into a single framework, advancing the system toward real-time applicability. Additionally, this work employs the publicly available CUBS database, which has been annotated to provide a valuable benchmark for future research. The proposed CADS demonstrates superior performance in noise reduction, segmentation accuracy, and classification tasks. Unlike most prior studies that focus only on segmentation or disease detection, the system not only distinguishes normal and abnormal CCA images but also categorizes plaque type. By unifying all these components, the framework emerges as a distinct and comprehensive solution for CCA detection and grading.

The innovations and contributions of this work are as follows: Unlike earlier approaches, this paper proposes a fully automated pipeline for carotid artery disease diagnosis and plaque grading. A novel two-level classification strategy combining ViT and SVM is introduced. In this framework, ViT distinguishes normal from abnormal cases, while SVM grades plaque severity, a hierarchical structure not present in earlier approaches. This multistage design achieves higher accuracy than conventional single-level classifiers. A CMF is employed to effectively denoise images while preserving vessel structures, and a threshold-based segmentation using Modified AOA ensures precise ROI extraction, outperforming other methods. The framework is validated on the CUBS database, demonstrating improved scalability and robustness compared to existing methods. To the best of the author’s knowledge, this is the first fully automated CADS framework designed for both diagnosis and plaque grading in carotid artery disease.

4.1 Details of the data set

The proposed CADS system was designed and tested using MATLAB 2022a running on an Intel Core i5 CPU @ 2.73 GHZ processor with 16GB RAM. Performance of the proposed CADS was tested on the CUBS database [31]. The database consisted of 2176 images, which included both normal and abnormal images. A broadband L12-5 MHz linear array transducer was used with a Philips (ATL) HDI 5000 duplex scanner for all scans. For every topic, numerous images were taken, and the operator who captured the image selected the finest ones.

All acquisitions used the same set of US characteristics, such as high frame rate and low persistence. A trained operator used a MyLab25 equipment with a LA523 4-13 MHz linear array transducer to acquire all scans. It used both digital and manual measuring techniques for all 2175 images. From this database, a total of 950 CCA images, which include 455 normal images and 495 abnormal images were chosen for validation. The effectiveness of the proposed CADS system was assessed by a 10-fold cross-validation method.

4.2 Evaluation metrics

Peak-Signal-to-Noise Ratio (PSNR) and Correlation Coefficient (CC), which measure how comparable noisy and denoised images are, have been used to assess the denoising performance of the suggested CMF.

$PSNR=10.lo{{g}_{10}}\left( \frac{ma{{x}^{2}}}{MSE} \right)$         (17)

$\text{MSE}=\frac{1}{MXN}\underset{i=1}{\overset{M}{\mathop \sum }}\,\underset{j=1}{\overset{N}{\mathop \sum }}\,{{\left[ I\left( i,j \right)-K\left( i,j \right) \right]}^{2}}$         (18)

$\text{NCC}=\frac{\mathop{\sum }_{i=1}^{M}\mathop{\sum }_{j=1}^{N}\left[ I\left( i,j \right)-\bar{I} \right]\left[ K\left( i,j \right)-\bar{K} \right]}{\sqrt{\mathop{\sum }_{i=1}^{M}\mathop{\sum }_{j=1}^{N}{{\left[ I\left( i,j \right)-\bar{I} \right]}^{2}}\mathop{\sum }_{i=1}^{M}\mathop{\sum }_{j=1}^{N}{{\left[ K\left( i,j \right)-\bar{K} \right]}^{2}}}}$         (19)

The degree of similarity between the suggested approaches and the ground truth has been measured using the Jaccard Similarity Coefficient (JSC) and Dice Coefficient (DC) to evaluate the ROI segmentation performance [32]:

$\text{JSC}\left( \text{P},\text{Q} \right)=\frac{\left| P\cap Q \right|}{\left| P\cup Q \right|}$         (20)

$\text{DC}\left( \text{P},\text{Q} \right)=\frac{2\left| P\cap Q \right|}{\left| P \right|+\left| Q \right|}$        (21)

Matthew's Correlation Coefficient (MCC), Balanced Accuracy (BA), Negative Predictive Value (NPV), F1-score, sensitivity, specificity, accuracy, and precision have all been used to assess the classification performance.

$\text{Accuracy }\!\!~\!\!\text{ }=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}}$       (22)

$\text{Sensitivity}/\text{Recall}=\frac{\text{TP}}{\text{TP}+\text{FN}}$         (23)

$\text{Specificity }\!\!~\!\!\text{ }=\frac{\text{TN}}{\text{TN}+\text{FP}}$       (24)

$\text{BA}=\frac{Sensitivity+Specificity}{2}$       (25)

$\text{MCC}=\frac{\text{TPXTN}-\text{FPXFN}}{\sqrt{\left( \text{TP}+\text{FP} \right)\left( \text{TP}+\text{FN} \right)\left( \text{TN}+\text{FP} \right)\left( \text{TN}+\text{FN} \right)}}$        (26)

Precision $=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$        (27)

$F 1$ score $=2 \times \frac{\text { Precision } \times \text { Recall }}{\text { Precision }+ \text { Recall }}$          (28)

$\text{Specificity }\!\!~\!\!\text{ }=\frac{\text{TN}}{\text{TN}+\text{FN}}$      (29)

where, TP- True Positive, TN- True Negative, FP- False Positive, and FN- False Negative.

4.3 Simulation results

4.3.1 Denoising performance analysis

In this work, CMF with a radius of 7 was employed to remove noise from the US images. Effectiveness of the CMF was assessed and compared with three different types of filters such as MF with a kernel size of 3 × 3, Weiner Filter (WF) with a kernel size of 3 × 3, and Gaussian Filter (GF). Quantitative analysis of these filters is reflected in Table 1. This analysis is based on two metrics: PSNR and NCC. As observed in Table 1, the CMF demonstrated the best performance with the highest PSNR and NCC, indicating it is highly effective in preserving both the quality and the structural details of the medical images. This superior performance of the CMF was attributed to its ability to use a circular neighborhood for median calculation, which is particularly beneficial in handling complex structures in the medical images without introducing artifacts. The MF filter operated on a 3 × 3 square neighborhood, did not capture the spatial variations as effectively as the CMF. This limitation resulted in lower PSNR and NCC values. While MF was good at noise reduction, it neither did nor preserved details as well.

Table 1. Performance comparison of various denoising filters

The GF applied a weighted average based on a Gaussian distribution, which provided moderate performance. The output was blurred, leading to lower PSNR and MCC. This blurring effect contributed to the loss of critical details in medical images. The WF reduced noise based on local statistics. It showed the least performance with the lowest PSNR and NCC. Unlike MF, which used a square-shaped neighborhood, the CMF used a circular neighborhood. This shape was more natural and better aligned with the way features are distributed in images.

(a) Noisy image

(b) Median filter output

(c) Weiner filter output

(d) Gaussian filter output

(e) CMF output

Figure 6. Visual analysis of various denoising filters

The use of a larger radius allowed the CMF to consider a broader range of pixel values, making it more robust against high levels of noise while still preserving important features. This combination of a circular neighborhood and a larger radius enabled the CMF to outperform other filters, making it an ideal choice for the denoising of US medical images. Qualitative analysis of these filter outcomes is shown in Figure 6. The CMF processed images retained more of the original texture and structural details, while the images processed by MF, GF, and WF exhibited varying degrees of blurring and detail loss.

4.3.2 Segmentation performance analysis

After denoising, the proposed segmentation algorithm has been applied including edge detection, morphological operation, and Modified AOA optimized thresholding for isolating the affected region from its background. In Modified AOA, $\propto $ regulates the rate of transition from exploration to exploitation. A higher $\propto $ slows down exploitation, whereas a lower $\propto $ accelerates convergence but may risk premature local optimal. After testing, $\propto $ was set to 0.5, which provided an optimal trade-off between convergence and segmentation accuracy. The parameter k adjusts the inertia weight, which controls the step size during exploitation. A larger k increases exploration in early iterations, while a smaller k promotes fine-tuning in later iterations. The value of k was set to 0.8, which improved stability and prevented premature convergence. To validate these choices, a sensitivity analysis was conducted by varying $\propto $ and k across different ranges. Table 2 shows the performance, and the analysis shows that $\propto $ = 0.5 and k=0.8 yielded the best balance between segmentation accuracy and efficiency.

Table 2. Sensitivity analysis

Figure 7. Comparative analysis of various segmentation algorithms

To demonstrate how the proposed technique is superior, segmentation performance was computed and compared against some earlier methods reported in the literature. All the past approaches were evaluated using the same images to ensure a realistic comparison. Figure 7 shows the segmentation performance comparison with earlier approaches. It is quite clear from Figure 7 that the proposed methods attained the highest JSC value of 0.836 and DC value of 0.911, indicating the best overall segmentation accuracy and overlap with the ground truth. The proposed Modified AOA optimized threshold incorporated an advanced optimization technique that significantly enhanced the segmentation accuracy. The combination of Affinity Propagation followed by Density Based Spatial Clustering of Application with Noise (AB+DBSCAN) provided better performance than other methods, however, lower than the proposed method.

Mohammed et al. [1] applied Deepnet to separate the ROI from its background. The net attained JSC of 0.756 and DC of 0.861. However, it might not be as optimized as the proposed Modified AOA method. Latha et al. [17] utilized the AB+DBSCAN method and reported JSC of 0.764 and DC of 0.865. Active oblongs were used for image segmentation by Dhupia et al. [33]. This method yielded JSC and DC of 0.721 and 0.836, respectively. The active oblongs model may struggle with highly complex images, which can affect segmentation accuracy. Yuan et al. [34] designed CMS-Net for lumen CCA segmentation. This net has the lowest JSC of 0.702 and DC of 0.825 among the compared methods. While CMS-Net might be effective and efficient in some scenarios, it could be less effective in handling variations in lumen shape, leading to lower performance compared to other sophisticated methods. From the analysis, it is evident that the proposed method outperformed all other methods due to its optimization process that enhances segmentation accuracy.

4.3.3 Classification performance analysis

The proposed CADS has utilized ViT to categorize the CCA US images into normal (no plaque) or abnormal (plaque). A 10-fold cross-validation strategy was employed to evaluate generalization and minimize the risk of overfitting. To ensure that each sample contributed to both training and validation across runs, nine folds were employed for training and one fold for validation in each iteration. Training accuracy and loss plots are shown in Figure 8. The graph demonstrates stable convergence with accuracy approaching 100% and loss decreases smoothly toward zero. The consistency of results across folds confirms robustness and indicates that the model generalizes well without significant overfitting. The categorization performance has been evaluated using a number of assessment parameters of the model and these results have been contrasted against earlier methods, including ResNet-50 [20], MobileNet [35], CNN [36], AlexNet [37], and UNet++ [38]. Table 3 presents a comparative analysis of the proposed method against these alternatives.

Figure 8. Training process (a) Training accuracy (b) Training loss

Table 3. Comparative analysis of different methods for Common Carotid Artery disease classification

On observing classification accuracy, the proposed CADS achieved the highest accuracy, indicating its superior capability to correctly diagnose CCA. While MobileNet performed better than AlexNet and CNN, its classification accuracy was lower than that of the proposed CADS. The MobileNet, CNN, and AlexNet reported specificity values of 82.61%, 69.57%, and 71.74%, respectively. With the integration of more advanced architectures, ResNet-50 and UNet++ showed significant improvements across all metrics, achieving accuracies of 88.45% and 90.12%, respectively. However, the proposed model posed the highest specificity of 91.3%, indicating it is best at detecting TNs. Similarly, the proposed model excelled in recall, showing it accurately detected most of the positive cases. Precision was highest for the proposed model, indicating that it predicts most of the positive cases. Although AlexNet was also strong in this metric, it fell short compared to the CADS.

NPV value represents the ability of a classifier to correctly predict true negatives. The proposed CADS achieved the highest NPV, indicating it is very effective at predicting TNs. BA accounts for both sensitivity and specificity. The F1-score balances precision and sensitivity. The CADS gave the highest F1-score, indicating a well-balanced performance in precision and sensitivity. Moreover, the proposed CADS achieved the highest MCC, reflecting the quality of its classification. MCC value showed a strong correlation between predicted and actual classes. MobileNet exhibited moderate MCC, AlexNet and CNN had lower values.

Although MobileNet showed better accuracy and recall than AlexNet and CNN, it still produced more FPs compared to the CADS. Basic CNN can be easier to implement. This model yielded lower performance across all metrics compared to other models. Hence, less effective at diagnosing carotid artery disease. AlexNet performed better than CNN and offered reasonable accuracy, sensitivity, and precision. From the detailed analysis, it is evident that the proposed CADS demonstrated superior performance in diagnosing carotid artery disease compared to past methods.

Figure 9. Receiver operating characteristic curve for common carotid artery diagnosis

A binary classifier system's ability to diagnose problems when its discriminating threshold changes is shown graphically by the Receiver operating characteristic (ROC) curve. The True Positive Rate (TPR) is plotted against the False Positive Rate (FPR) across multiple threshold values to get this curve. A model with an Area Under Curve (AUC) of 0 has no discrimination ability, while an AUC of 1 represents perfect discrimination. From Figure 9, it is observed that MobileNet [35] showed a good ability to distinguish between classes. The ROC curve for MobileNet showed that the TPR is significantly higher than the FPR across most threshold settings, indicating that it effectively identifies TP while minimizing FPs. The CNN [36] model showed moderate performance. While it achieved an acceptable ability to distinguish between classes, its performance was lower than that of MobileNet and CADS. AlexNet [37] demonstrated fair performance, slightly better than the CNN but still lower than CADS.

The proposed model exhibited an excellent performance, significantly surpassing all other models in its ability to distinguish between classes. The ROC curve for the proposed model indicates a much higher TPR relative to FPR across a broad range of thresholds, reflecting its superior capability in correctly classifying instances. This outstanding performance highlights the CADS’s robustness and effectiveness as a classifier, making it the most reliable among the models evaluated.

4.3.4 Plaque grading performance analysis

The proposed system used SVM to grade the plaque in abnormal cases. The efficacy of the suggested CADS plaque grading is seen in Figure 10. The model gave an accuracy of 86.32% which means that the CADS correctly categorizes classes. The model was quite effective at correctly detecting negative cases with a specificity of 86.96%. An 85.71% recall showed the model was good at identifying positive cases. The model had a high probability of identifying positive cases are TP with a precision of 87.5%. An NPV of 85.11% indicates that most of the negative predictions are accurate. A BA of 86.33% indicated robust performance. F1-score of 86.6% reflected a good balance between precision and recall, representing the CADS’s reliability in positive class prediction. An MCC of 72.6% represented overall good performance. The results proved that the model is well-suited for the task of plaque classification with high accuracy, balanced specificity and sensitivity, and strong overall reliability. ROC curve for plaque classification is depicted in Figure 11.

Figure 10. Plaque classification performance

Figure 11. Receiver operating characteristic (ROC) curve analysis for plaque grading

4.4 Discussion

In this paper, a CADS to diagnose CCA disease and grade plaque is proposed and implemented. The importance of this system lies in its potential to address carotid artery structure, which is characterized by carotid arterial narrowing. The disease is particularly insidious because it develops gradually over time, often starting in childhood and slowly progresses without noticeable symptoms. Preliminary identification is important to prevent the progression of atherosclerosis. By implementing this CAD, the purpose of this study is to provide a timely and accurate clinical tool that can significantly reduce the risk of stroke by detecting and classifying arterial plaque in its development.

The most significant contribution of this investigation is the introduction of a new framework that encompasses five distinct phases: denoising, lumen segmentation, classification of CCA as normal and abnormal, and further plaque classification. This structured approach not only enhances the accuracy of the CCA diagnosis but also addresses the gap in the literature, where a complete system for CCA diagnosis and plaque classification has yet to be fully proposed and reported. Each phase of the framework plays a vital role in ensuring that the CADS operates with high accuracy. For instance, the denoising phase is essential for improving the quality of the US images, thereby facilitating more accurate segmentation and classifications. The segmentation phase focuses on detecting the lumen, which is crucial for assessing the degree of narrowing. By incorporating these elements into a single system, this study pioneers a comprehensive solution to CCA diagnosis that is not only innovative but highly practical for clinical use.

A primary feature of CADS is the ability to combine multiple processes, including noise removal, segmentation and classification in the same structure. It addresses an important challenge in the integration field, where each of these processes is usually handled separately, often requiring different devices. By uniting these steps, the proposal system streamlines the clinical process, making it more efficient and less resource-intensive. The classification phase within this structure allows for the exclusion of normal images from the plaque classification process. This step is crucial as it reduces the computational load and speeds up the overall diagnosis, making the system more suitable for real-time deployment. 

The comparative analysis in Table 3 clearly shows that the proposed CADS attained excellent performance across various metrics compared to other methods. The key observations are as follows:

• The proposed CADS achieved an accuracy of 92.63%, surpassing the Chatterjee et al. [35] method’s accuracy of 82.11%, outperforming the Abd-Ellah et al. [37] method with 74.74% accuracy, and significantly exceeding the Varma et al. [36] method, which reported an accuracy of 70.53%. This significant improvement demonstrates the effectiveness with which the proposed CADS performs CCA image classification.
• The proposed CADS attained a specificity of 91.3%, significantly outperforming the specificity of 82.61% by Chatterjee et al. [35], 69.57% by Varma et al. [36] and 71.74% specificity by Abd-Ellah et al. [37]. This marked improvement underscores the performance of the CADS.
• The proposed CADS reported a sensitivity of 93.88%, surpassing the sensitivity of 81.63% attained by Chatterjee et al. [35], far exceeding the 71.43% sensitivity of Varma et al. [36] and outperforming the 77.55% reported by Abd-Ellah et al. [37].
• In terms of precision, the CADS reached 92%, a significant improvement over the 83.33% precision achieved by Chatterjee et al. [35], 71.43% by Varma et al. [36] and 74.51% reported by Abd-Ellah et al. [37]. This substantial gain highlights the robustness of CADS.
• Additionally, the proposed CADS recorded a BA of 92.58%, while Chatterjee et al. [35] reported 82.12%, Varma et al. [36] attained 70.49%, and Abd-Ellah et al. [37] achieved 74.59%.

The potential for real-time deployment of the CCA diagnosis system is one of the most significant advantages. Real-time processing is particularly valuable for radiologists, who require quick and accurate results to make informed decisions about patient care. By minimizing processing time, the system ensures that the diagnosis can be made promptly, which is essential in acute medical situations. The proposed model required approximately 53 seconds for training and 1.01 seconds for testing. These runtimes demonstrate that the model can be executed efficiently on standard hardware without the need for specialized computational resources. The outcomes indicate that the approach is computationally feasible for practical applications. The CADS system not only improves the speed and accuracy of the CCA diagnosis but also offers a scalable solution that can be adapted for use in a variety of clinical settings. As carotid artery disease remains a leading cause of stroke, the implementation of such advanced diagnostic tools is essential for improving patient outcomes and preventing severe complications associated with the disease.

The model was evaluated under standard experimental conditions without testing across varying noise levels, alternative data distributions, or imaging devices. Future studies will address this limitation by incorporating robustness experiments with noisy or low-quality images as well as validation on cross-device and multi-institutional datasets to ensure stability and broader applicability.