Intelligent detection method for carotid atherosclerotic plaques based on ultrasound imaging

Abstract

BACKGROUND:

The formation of carotid atherosclerotic plaques (CAP) is a major contributor to cardiovascular and cerebrovascular diseases. Early detection and accurate identification of plaques are crucial for reducing the risk of these conditions. However, traditional CAP detection methods rely heavily on manual analysis by clinical experts and specific imaging equipment, which limits their feasibility for large-scale screening and automated processing. Therefore, developing efficient and automated methods for CAP detection holds significant clinical value.

METHODS:

This paper presents an intelligent CAP detection approach based on ultrasound imaging. By extracting texture and morphological features from the images, such as gray-level co-occurrence matrix, gray-level run length matrix, invariant moments, and fractal dimension, we construct a machine learning model using random forest (RF) and other algorithms to achieve accurate and efficient CAP detection. We evaluated the model’s performance on a dataset of 2,786 ultrasound images acquired from various devices, employing five-fold cross-validation.

RESULTS:

The experimental results indicate that the proposed machine learning model achieves high CAP detection accuracy across multiple classification metrics. The optimal RF model demonstrated an accuracy of 80%, a precision of 82%, a recall of 89%, a specificity of 75%, and an F1 score of 85%. Furthermore, feature importance analysis allows us to identify the ultrasound image features most relevant to CAP detection, enhancing the interpretability of our model.

CONCLUSION:

The machine learning model developed in this study effectively supports automated CAP detection and provides a reliable foundation for intelligent diagnosis using ultrasound imaging.

Introduction

According to statistics, by 2020, approximately two billion people worldwide were affected by carotid atherosclerosis.[1] The formation of carotid atherosclerotic plaques (CAP) is a significant risk factor for cardiovascular and cerebrovascular diseases, closely linked to life-threatening conditions such as stroke and myocardial infarction. Research indicates that within 10 years, patients with CAP have a 4% incidence rate of myocardial infarction and a 40% increased risk of stroke.[2] The development of CAP leads to thickening of the carotid artery wall and narrowing of the lumen, which impairs blood flow and increases the risks of thrombosis and embolism.[3] The clinical management of CAP depends on their hemodynamic significance and associated risk profiles. Current guidelines emphasize that asymptomatic plaques with low-to-moderate stenosis (typically < 70%) or stable morphology are often managed conservatively through lifestyle modifications and pharmacological therapy (e.g., statins and antiplatelet agents).[4] In contrast, high-risk plaques—characterized by severe stenosis (> 70%), ulceration, intraplaque hemorrhage, or symptomatic presentation—may necessitate surgical (carotid endarterectomy) or endovascular interventions (carotid stenting). However, a critical challenge lies in accurately identifying high-risk plaques during early screening, as ultrasound-based detection often serves as the first-line diagnostic tool. While our study focuses on plaque detection rather than treatment stratification, the proposed automated method enhances the efficiency of preliminary risk assessment, enabling timely referral for advanced imaging when warranted. Thus, early detection and accurate assessment of CAP are of critical clinical importance for preventing cardiovascular diseases.[4] Traditional CAP detection relies heavily on manual analysis by clinicians[5,6] and specialized imaging equipment, including ultrasound,[7,8,9] computed tomography,[10] and magnetic resonance imaging.[11] These methods demand extensive professional expertise and are often highly subjective, time-consuming, and cost-prohibitive.[12] As a result, they are inefficient for large-scale screening and unable to meet the increasing demand for fast, accurate, and automated CAP detection.

With advancements in artificial intelligence and medical imaging analysis,[13,14,15] intelligent detection methods have emerged as essential tools for enhancing clinical detection efficiency.[16,17,18] For the effective and accurate detection of CAP, there is an urgent need for an automated machine learning approach with strong interpretability that can replace traditional expert-dependent methods. By leveraging the textural and morphological features in ultrasound images, this approach can effectively capture the structural information of CAP, enabling precise classification and identification. It is important to note that patients undergoing intervention typically require supplementary imaging, such as CTA or magnetic resonance angiography, to evaluate aortic arch anatomy, plaque calcification, and collateral circulation—factors critical for surgical or endovascular planning. Ultrasound, while noninvasive and cost-effective, lacks the spatial resolution to comprehensively assess these anatomical details. Nevertheless, the proposed intelligent detection method aims to streamline the preliminary screening process, reducing reliance on subjective manual analysis and facilitating earlier identification of high-risk cases. By improving the accuracy and accessibility of initial CAP detection, our approach complements—rather than replaces—the role of CTA in treatment planning, ensuring that advanced imaging resources are allocated efficiently.

Thus, the primary objective of this study is to evaluate and compare the performance of machine learning models—including random forest (RF), decision tree, gradient boosting, and others—in detecting CAP using ultrasound imaging. By leveraging texture and morphological features (such as gray-level co-occurrence matrix [GLCM],[19] gray-level run length matrix [GLRLM],[20] invariant moments,[21] and fractal dimension [FD][22]), we aim to develop an automated, interpretable framework that enhances the accuracy and efficiency of CAP detection, thereby addressing the limitations of manual, expert-dependent methods.

The main contributions of this study are threefold: first, we propose a machine learning framework that integrates diverse ultrasound image features (GLCM, GLRLM, invariant moments, FD) to improve CAP detection accuracy. Second, our approach emphasizes model interpretability through feature importance analysis, enabling clinicians to understand the basis of automated decisions. Third, we validate the framework on a multi-device dataset of 2,786 images, ensuring robustness across imaging conditions.

Methods

Data source

This study utilizes a dataset of carotid ultrasound images collected from 147 outpatient patients by the Functional Department of Yichun People’s Hospital in Jiangxi Province. The dataset construction process followed a systematic workflow [Figure 1]. The process includes patient enrolment, standardized image acquisition across multiple devices, quality control (no exclusions), expert annotation of plaque regions, and dataset finalization. The dataset acquired using various devices, including the Philips EPIQ 7C (probe L12-5, MI 1.3, TIS 0.3), Philips IE33 (probe L9-3, MI 0.9, TIS 0.3), Mindray DC–75 Exp (probe L12-3E, MI 1.4, TIS 0.3), and Aloka Prosound F75 (probe L12-5, MI 1.4, TIS 0.4). This diversity in equipment ensures the model’s applicability across varied data sources. All ultrasound examinations followed a standardized protocol aligned with the American Institute of Ultrasound in Medicine guidelines. Radiologists obtained longitudinal and transverse B-mode grayscale images of the carotid bifurcation and proximal internal carotid artery, with patients in a supine position and the neck rotated 45° contralaterally. The scanning depth, gain, and dynamic range were optimized to maximize plaque visualization while minimizing artifacts. Images were excluded only if incomplete (e.g., partial probe contact, loss of anatomical landmarks) or corrupted during acquisition. All 2,786 images were retained after quality control, ensuring representation of real-world clinical variability. Expert annotation and consensus resolved discrepancies, guaranteeing reliable labels for model training.

Figure 1.

Dataset construction workflow. Visual representation of the systematic workflow for dataset construction, from patient enrolment to finalization of the 2786-image dataset

The 147 outpatient patients (mean age: 65.3 ± 9.7 years; 62% male) presented with symptoms or risk factors (e.g., hypertension, dyslipidemia, and smoking history) warranting carotid ultrasound screening per clinical guidelines. The dataset includes both asymptomatic patients (n = 89) and those with transient ischemic attack or stroke history (n = 58), reflecting a representative spectrum of CAP presentations. The high image-to-patient ratio (≈19 images per patient) ensures comprehensive spatial sampling of plaque morphology across multiple carotid segments, which is critical for training robust machine learning models.

Data preprocessing

To mitigate the effects of varying equipment and imaging conditions on image quality, a systematic preprocessing pipeline was applied, as shown in Figure 2. First, noise reduction was performed using median filtering and nonlocal means filtering[23] to remove speckle noise while preserving edge details. Next, contrast stretching and histogram equalization[24] were applied to enhance image contrast, making CAP regions clearer and improving the accuracy of subsequent feature extraction. Following this, images were cropped and standardized to a uniform size, focusing on the carotid region and removing irrelevant background, thus ensuring input data consistency. Finally, all images were subjected to grayscale normalization, standardizing intensity values across the dataset and enhancing the model’s robustness to intensity variations between images.

Figure 2.

Schematic diagram of ultrasound image preprocessing process

Feature extraction

To enhance the accuracy of CAP detection, four types of features were extracted from the ultrasound images, capturing both texture and morphological information. All the extracted features are listed in Figure 3.

Figure 3.

Schematic diagram of ultrasound image preprocessing process

Gray-level co-occurrence matrix features

GLCM features capture the texture information within an image. Given the gray levels of an image, GLCM describes the spatial relationships between pairs of gray values, defined as the frequency with which gray levels i and j occur at a specified distance d and orientation θ. Let P (i, j | d, θ) represent the value at position (i, j) in the GLCM; the following features were calculated in this study:

Entropy: Measures the complexity of the texture.

Energy: Assesses the smoothness of the ultrasound image.

Contrast: Describes the difference between gray levels.

Homogeneity: Reflects the similarity of pixels.

These features were computed at different orientations (e.g., 0°, 45°, 90°, and 135°) and distances, allowing the model to capture detailed textural characteristics in CAP regions.

Gray-level run length matrix features

GLRLM features describe the distribution of consecutive pixels with the same gray level, particularly useful for capturing linear structures within the image.[20] Let R (i, j) represent the occurrence count of consecutive pixels with gray level i and run length j; the following GLRLM features were extracted:

Short run emphasis (SRE):

Long run emphasis (LRE):

Gray level nonuniformity (GLN):

Run length nonuniformity (RLN):

Run percentage (RP): Assesses the flatness of the image.

These features provide detailed information about the surface texture of CAP regions, enriching the model’s ability to differentiate plaques based on texture.

Invariant moment features

Invariant moment features possess rotation, scale, and transformation invariance,[21] making them suitable for describing the shape of plaques. The seven Hu moments φ₁ through φ₇ were derived from the second- and third-order moments of the image. For each ultrasound image, the central moments μ_pq were computed:

The normalized central moments η_pq were then calculated:

Using η_pq, Hu invariant moments were constructed, for example, the first Hu moment φ₁ is defined as:

These features characterize the geometric properties of the plaques, providing essential morphological information for the model.

Fractal dimension features

FD features capture the complexity and boundary roughness of CAP regions.[22] The commonly used box-counting method estimates FD by counting the number of boxes N(ε) needed to cover the object at various scales:

FD features capture fine details and irregularities along the plaque boundaries, aiding in the identification of complex lesion regions.

Model construction

After completing feature extraction, multiple machine learning models were constructed for CAP detection, as shown in Figure 3. The RF model was recommended due to its robust classification capability and stability in multi-dimensional feature spaces. As shown in Figure 4, RF constructs multiple decision trees during training and combines their outputs (usually by averaging for regression or majority voting for classification) to improve predictive accuracy and reduce overfitting. In addition, other models, including Decision Tree, Gradient Boosting, AdaBoost, Logistic Regression, Linear Support Vector Machines (SVM), and radial basis function (RBF) SVM, were implemented as benchmarks to comprehensively assess performance across models for this task. The experiments were conducted on a server equipped with an Intel Xeon processor, 32 GB of memory, and an NVIDIA 2070 GPU to accelerate the training process. All models were implemented in a python environment using the Scikit-learn library.

Figure 4.

The construction principle diagram of random forest

To ensure robust results, model hyperparameters were optimized. For the RF model, the number of trees was set to 100, and the maximum depth was fine-tuned through cross-validation. Hyperparameters for other comparison models were also optimized using grid search to achieve the best performance. Five-fold cross-validation was employed throughout to evaluate the robustness and generalizability of each model. Besides, to comprehensively evaluate detection performance, standard classification metrics were used, including accuracy, precision, recall, F1 score, and area under the curve (AUC). All implementation code for the models is publicly available in a GitHub repository (https://github.com/Ma-Chenbin/CAPD_ML).

Model interpretability analysis

To improve model interpretability, feature importance analysis was conducted to identify the features most influential in CAP detection. Feature importance analysis provided weights for each feature in the model’s decision-making process, revealing the roles of both texture and morphological features in CAP detection. In addition, visualization of key features was performed to illustrate their contributions to detection, enabling clinicians to better understand the model’s decision basis and further enhancing its clinical applicability.

Clinical trial registry

Not applicable since no clinical trials were involved.

Results

Detection performance comparison

Using a dataset of 2,786 ultrasound images, we evaluated and compared the primary metrics of each model using five-fold cross-validation. Detailed results are presented in Table 1, with a corresponding boxplot visualization shown in Figure 5. Statistical significance between the RF and baseline models was assessed using Wilcoxon signed-rank tests [Table 2].

Table 1.

Comparison of five-fold cross validation results of machine learning models (mean±standard deviation)

Models	Accuracy	Precision	Recall (sensitivity)	Specificity	F1 score	AUC
RF	0.80±0.01	0.82±0.03	0.89±0.01	0.75±0.02	0.85±0.01	0.87±0.01
Decision tree	0.72±0.02	0.78±0.03	0.79±0.01	0.71±0.03	0.78±0.02	0.70±0.02
Gradient boosting	0.78±0.01	0.78±0.03	0.90±0.01	0.69±0.02	0.84±0.01	0.84±0.01
AdaBoost	0.76±0.01	0.78±0.03	0.88±0.00	0.66±0.02	0.83±0.01	0.81±0.02
Logistic regression	0.67±0.02	0.69±0.03	0.90±0.01	0.58±0.03	0.78±0.02	0.68±0.02
Linear SVM	0.67±0.02	0.68±0.04	0.90±0.01	0.55±0.03	0.78±0.02	0.68±0.02
RBF SVM	0.75±0.02	0.75±0.03	0.92±0.00	0.56±0.04	0.83±0.02	0.78±0.01

RF: Random forest, SVM: Support vector machines, RBF: Radial basis function, AUC: Area under the curve

Figure 5.

Box plot visualization of the five-fold cross validation of the machine learning model regarding (a) Accuracy, (b) Precision, (c) Recall (sensitivity), and (d) F1 score. RBF: Radial basis function, SVM: Support vector machines

Table 2.

Statistical comparison between models (P)

Model pair	Accuracy	Precision	Recall	Specificity
RF versus gradient boosting	< 0.001	0.003	0.12	< 0.001
RF versus RBF SVM	< 0.001	< 0.001	0.002	< 0.001
RF versus decision tree	< 0.001	0.001	< 0.001	< 0.001
RF versus AdaBoost	< 0.001	0.002	0.05	< 0.001
RF versus logistic regression	< 0.001	< 0.001	0.2	< 0.001
RF versus linear SVM	< 0.001	< 0.001	0.3	< 0.001

Random forest, SVM: Support vector machines, RBF: Radial basis function, AUC: Area under the curve

The RF model demonstrated robust performance in carotid CAP detection, achieving an accuracy of 80.40% ± 1.00%, significantly surpassing gradient boosting (77.89% ± 1.17%, P < 0.001) and decision tree (72.18% ± 2.00%, P < 0.001). Its precision of 82.00% ± 3.00% was statistically superior to all baseline models (P ≤ 0.003), reflecting its ability to minimize false positives–a critical advantage in clinical settings where overdiagnosis carries substantial economic and psychological costs. While the RBF SVM model exhibited marginally higher sensitivity (92.00% ± 0.00% vs. RF’s 88.80% ± 1.00%), this difference was not clinically significant (P = 0.002), particularly when contrasted with RF’s markedly superior specificity (75.00% ± 2.00% vs. RBF SVM’s 56.00% ± 4.00%, P < 0.001). Models such as Logistic Regression and Linear SVM, despite achieving high sensitivity (90.00% ± 1.00%), suffered from poor specificity (58.00% ± 3.00% and 55.00% ± 3.00%, respectively), resulting in unacceptably high false-positive rates that would burden clinical workflows with unnecessary follow-up imaging.

The RF model’s balanced performance—combining high sensitivity (88.80% ± 1.00%) and specificity (75.00% ± 2.00%)—translates to practical clinical benefits. For instance, its specificity of 75% implies that 25% of nonplaque cases may still require confirmatory imaging, aligning with current practices where ultrasound serves as a preliminary screening tool before advanced modalities like CTA. This balance reduces both missed diagnoses and overreferral risks, unlike models such as Linear SVM, which would erroneously flag 45% of healthy cases for further testing. Statistical comparisons further underscored RF’s robustness: it outperformed all models in accuracy (P < 0.001), with the largest improvement over logistic regression (13.37% absolute gain), and exhibited superior specificity across all benchmarks (P < 0.001). While its sensitivity trailed RBF SVM by 3.20%, this minor gap is outweighed by RF’s 19.00% higher specificity, ensuring fewer false alarms without compromising detection efficacy.

Figure 6 shows the ROC curves for the top three machine learning models. Among them, RF [Figure 6a] demonstrates the best performance, with an average AUC of 0.88, indicating strong discriminative ability and stability. Gradient Boosting [Figure 6b] follows with an average AUC of 0.84, slightly lower but still relatively consistent. AdaBoost [Figure 6c] has the lowest average AUC of 0.80, with noticeable variability across folds. Overall, RF is the most suitable model for applications requiring high accuracy.

Figure 6.

Receiver operating characteristic curve of the five-fold cross validation of the top-3 machine learning model: (a) Random forest, (b) Gradient Boosting, and (c) AdaBoost. ROC: Receiver operating characteristic, AUC: Area under the curve

Overall, the RF model’s consistently high metrics across precision, recall, specificity, F1 score, and AUC demonstrate its adaptability and reliability in CAP detection. The balance it achieves between high precision and recall enhances diagnostic accuracy while minimizing both false positives and false negatives, confirming its practical value in clinical diagnosis.

Feature importance analysis

Additionally, Table 3 presents each model’s feature importance scores, and normalized scores are visualized in Figure 7. The feature importance analysis highlights significant differences in each model’s reliance on specific features, revealing key features’ roles and contributions within the classification task.

Table 3.

Importance scores of ultrasound image features by machine learning models

Features	RF	Gradient boosting	AdaBoost	RBF SVM	Decision tree	Logistic regression	Linear SVM
HuMoment_2	0.171	0.309	0.162	0.003	0.217	0.469	0.196
Long run high gray level emphasis	0.118	0.136	0.114	0.057	0.145	0.169	0.075
HuMoment_1	0.114	0.124	0.134	0.006	0.093	0.431	0.182
GLN	0.114	0.076	0.142	0.041	0.099	0.659	0.276
LRE	0.101	0.059	0.104	0.013	0.083	0.448	0.193
RLN	0.100	0.084	0.086	0.013	0.105	0.505	0.218
Contrast	0.097	0.084	0.076	0.047	0.100	0.699	0.378
FD	0.096	0.075	0.118	0.024	0.109	0.335	0.147
Idm	0.089	0.053	0.064	0.047	0.049	0.826	0.435

SVM: Support vector machines, RBF: Radial basis function, LRE: Long run emphasis, GLN: Gray level nonuniformity, RLN: Run length nonuniformity, FD: Fractal dimension, RF: Random forest

Figure 7.

Histogram visualization of the normalized importance scores of ultrasound image features by machine learning models. RBF: Radial basis function, SVM: Support vector machines

The results show that the Hu moment features are highly influential in several models, particularly in Gradient Boosting (0.309), Logistic Regression (0.469), and Decision Tree (0.217), where HuMoment_2 stands out among all features. This finding highlights the importance of Hu moments for capturing the geometric characteristics of CAP, with rotation and scale invariance making these features well-suited for CAP detection across various positions and scales. HuMoment_1 also shows high weight in Logistic Regression (0.431), further supporting Hu moments’ role in capturing shape complexity. While the reliance on HuMoment_1 is somewhat lower in gradient boosting and RF (0.124 and 0.114, respectively), its role in classification remains significant.

GLRLM features, specifically Long Run High Gray Level Emphasis and GLN, also show high importance across multiple models, indicating their value in representing image gray-level uniformity and run length. In the RF and Gradient Boosting models, the importance scores for Long Run High Gray Level Emphasis are 0.118 and 0.136, respectively, underscoring its role in tree-based models. GLN scores particularly high in logistic regression (0.659), suggesting that gray-level uniformity is essential in this model. In logistic regression and linear SVM, RLN holds importance scores of 0.505 and 0.218, indicating its role in distinguishing between diseased and normal tissue in linear models.

Within GLCM features, contrast has high importance in Logistic Regression (0.699) and Linear SVM (0.378), suggesting that gray-level contrast is crucial for these linear models in identifying plaque boundaries. The inverse difference moment feature is weighted at 0.826 in logistic regression and 0.435 in Linear SVM, showing its value in capturing texture smoothness and gray-level continuity, thus aiding linear models in recognizing uniform CAP regions and improving detection accuracy.

Finally, the FD feature is balanced across models, with moderate importance in RF, gradient boosting, and AdaBoost (0.096, 0.075, and 0.118, respectively). The FD is particularly effective in capturing the complexity of plaque boundaries, making it suitable for identifying irregular lesion areas and improving classification accuracy.

Therefore, feature analysis indicates that Hu moments play a crucial role in most models by capturing plaque shape characteristics. GLRLM and GLCM features enrich the model with texture information through gray-level distribution and contrast. Logistic regression’s strong reliance on GLCM and GLRLM features underscores their importance in linear models, while the FD enhances classification in tree-based models, especially in complex boundary regions. These findings validate the selected features’ effectiveness in CAP detection and underscore the method’s potential for model interpretability and clinical diagnostic support.

Discussion

Clinical utility

Traditional methods for detecting CAP typically rely on visual observation and measurement of ultrasound images.[8] This process often includes manual or semi-automated image segmentation and plaque analysis, largely dependent on clinicians’ experience and judgment to assess the severity of the disease by identifying arterial wall thickening and structural changes in the CAP.[5,6] Recent advances have been made with manually labeled and segmented plaque detection methods,[25,26] including techniques such as Canny edge detection[27] and convolutional neural networks based on supervised datasets[25,26] to localize CAP regions. However, these methods often suffer from limited accuracy and consistency due to noise and low contrast in ultrasound images, making manual detection highly subjective. Furthermore, traditional methods are not suitable for large-scale screening, as manual annotation is time-consuming and heavily dependent on the operator’s expertise, leading to inconsistent detection outcomes across different clinicians.[12]

To improve the efficiency and accuracy of CAP detection in ultrasound images, machine learning models have been widely applied in image analysis tasks. These methods can learn patterns from large volumes of annotated data, enabling automated image feature extraction and classification. In ultrasound image analysis, traditional machine learning algorithms such as SVM, RF, and Decision Tree have been applied to CAP detection.[28] These models typically rely on a set of informative features to perform classification within a multi-dimensional feature space. With the rise of interpretable machine learning, feature importance analysis now enables the identification of key features contributing to classification results, aiding in understanding the model’s decision-making process.[18,29,30] However, traditional machine learning approaches depend heavily on high-quality feature extraction, and the process of feature selection and model optimization often requires manual intervention, limiting the potential for full automation.[25,26,30]

Thus, while ultrasound image analysis has advanced through traditional and machine learning methods, challenges remain in terms of automation, detection accuracy, and interpretability. Building on prior work, this study combines texture and morphological features to construct an efficient machine learning model based on RF, demonstrating significant improvements in both detection accuracy and feature interpretability. Our study specifically addresses the automated detection of CAP presence in ultrasound images, which serves as a critical first step in clinical workflows. While the identification of plaque presence is essential for early risk stratification, it is distinct from assessing stenosis severity—a key determinant for intervention. Clinical guidelines emphasize that treatment decisions (e.g., surgical or endovascular intervention) require complementary evaluation of stenosis degree (> 70% for symptomatic or > 80% for high-risk asymptomatic cases), plaque morphology, and symptomatology. Current intervention criteria for CAP further emphasize hemodynamic significance (stenosis severity) and morphological instability (e.g., thin fibrous CAP). Current intervention criteria for CAP emphasize hemodynamic significance (stenosis severity) and morphological instability (thin fibrous CAP). For instance, the European Society for Vascular Surgery guidelines recommend revascularization for symptomatic patients with > 50% stenosis or asymptomatic patients with > 60% stenosis and high-risk features.

The RF model achieved high performance across key metrics, including accuracy (80.40%), precision (82.00%), and F1 score (85.22%), validating the effectiveness of using diverse texture and morphological features for CAP detection. By extracting features such as GLCM, GLRLM, Hu moments, and FD, the model captures the texture complexity and morphological characteristics of CAP regions, enhancing detection robustness. While the model detects plaque presence using grayscale texture and morphological features, emerging evidence suggests that such features correlate with plaque vulnerability. For instance, low GLCM entropy and high FD may indicate heterogeneous plaque composition—a marker of instability.[19] Similarly, invariant moments and RLN can capture irregular plaque boundaries associated with ulceration. Although the current study does not explicitly classify plaque subtypes, the extracted features provide a foundation for future risk stratification models. For example, GLRLM-based “LRE” may help distinguish calcified (stable) plaques from lipid-rich (unstable) ones, as longer runs correlate with homogeneous textures.

While ultrasound alone cannot fully characterize parameters like aortic arch anatomy or stenosis severity—parameters typically assessed via CTA for intervention planning—our automated detection model is designed for preliminary screening, not as a replacement for Doppler or angiography. By rapidly identifying plaque presence, the system prioritizes high-risk cases for subsequent Doppler-based velocity analysis or confirmatory imaging (e.g., CTA). This workflow aligns with clinical guidelines, where grayscale ultrasound precedes Doppler evaluation. Our automated detection method reduces variability in initial plaque identification, ensuring that high-risk candidates proceed efficiently to confirmatory imaging. For ulcerated plaques or dissections—conditions with limited ultrasound sensitivity—our method reduces diagnostic delays by flagging suspicious regions for angiographic validation. This workflow optimization is critical in resource-constrained settings, where manual screening delays may hinder timely intervention. In addition, our method provides high interpretability, allowing each feature’s contribution to be quantified through feature importance analysis, thus supporting understanding of the model’s decision-making process. Future iterations of this framework will integrate Doppler-based velocity analysis to estimate stenosis, but the current model’s focus on plaque detection aims to streamline preliminary screening, enabling timely referrals for advanced imaging when necessary.

Limitations

Despite the advantages in detection accuracy and feature interpretability, there are limitations in feature extraction and model applicability. First, the feature extraction process relies on hand-crafted texture and morphological features, which may not fully capture the complexity and diversity present in ultrasound images of different patients, potentially impacting the model’s generalizability. In addition, this study is based on static ultrasound images, leaving the applicability for dynamic scanning or real-time detection untested. Although five-fold cross-validation demonstrates model stability across images from different devices, further investigation on larger, more diverse datasets is needed to ensure the model’s robustness in broader clinical applications.

Future work

Future work could enhance this study by focusing on several areas. First, while our model detects plaque presence, it does not assess stenosis severity or morphological risk features (e.g., ulceration, lipid core), which are pivotal for treatment decisions. Future work will incorporate Doppler-derived velocity metrics and deep learning-based morphological analysis to address this gap. Second, the reliance on static images limits dynamic evaluation of plaque behaviour during real-time scanning. However, the framework’s high accuracy in plaque detection establishes a robust foundation for future extensions into stenosis grading and real-time applications. Further, applying the model to real-time ultrasound imaging and developing dynamic analysis techniques based on deep learning could enable real-time detection and assessment of CAP, further advancing the clinical utility of intelligent diagnostics.

Conclusion

This study presents an automated method for detecting CAP based on ultrasound images, leveraging texture and morphological features to construct an efficient machine learning model. By extracting features such as GLCM, GLRLM, invariant moments, and FD and integrating them with a RF model, this approach demonstrates superior performance in terms of detection accuracy and feature interpretability. Experimental results indicate that the proposed model achieves high accuracy and robustness across various classification metrics, validating its effectiveness for real-world CAP detection.

Author contributions

Concepts, design, definition of intellectual content and manuscript preparation: Shangming Zhu, Xunming Ji, Lishuang Guo, Xiaolian Qiu; Literature search, data acquisition and data analysis: Miaowen Jiang, Chongyuan He, Bo Sun, Jian Xu; Manuscript editing: Shangming Zhu, Lishuang Guo; Manuscript review: Jinlong Ma, Min He, Xunming Ji.

Ethical policy and institutional review board statement

This study adheres to the principles of the Declaration of Helsinki. The ethical approval for the study was obtained from Yichun People’s Hospital Medical Ethics Committee (No. 092, dated on 21^st March, 2024) and all participants provided informed consent as required.

Data availability statement

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Conflicts of interest

Prof. Xunming Ji is the Editor-in-Chief of Brain Circulation. The article was subject to the journal’s standard procedures, with peer review handled independently of this Editor and their research groups.

Funding Statement

This research was supported by National Natural Science Foundation of China (82027802), Science and Technology Innovation Service Capacity Building Project of Beijing Municipal Education Commission (11000023T000002157177). The funding bodies had no involvement in the design, analysis, or interpretation of the research.