A clinically deployable deep learning model for automated stroke risk stratification in carotid atherosclerotic plaque

Background: Carotid ultrasound is widely utilized for early risk screening of ischemic stroke. However, the accuracy and reproducibility of assessing plaque vulnerability-related features remain constrained by physicians' subjective interpretation, underscoring an urgent need to achieve precise and objective assessment of these features through intelligent quantification. Objective: This study aims to develop and compare deep learning (DL) and conventional machine learning (ML) models based on carotid plaque ultrasound images, so as to identify the optimal clinically applicable algorithm for precise plaque assessment and risk prediction. Methods: In this retrospective cohort study, 666 patient's carotid plaque ultrasound images (299 stroke patients; 367 non-stroke controls) collected between 2021 and 2025 were analyzed. Five convolutional neural networks (CNNs, e.g., ResNet-50) and two conventional machine learning (ML) classifiers [support vector machine (SVM), logistic regression (LR)] were trained on region-of-interest annotated plaque images using an 8:2 training-to-validation split. The area under the receiver operating characteristic curve (AUC) served as the primary performance metric, supplemented by accuracy, sensitivity, and specificity as secondary evaluation indices. The stroke risk prediction efficacy of the optimal DL model was subsequently compared with that of the ML models. Results: = 0.554). Critically, the best-performing DL model (ResNet-50) exhibited a 9.7% improvement in AUC over the top ML model (0.982 vs. 0.885), with consistently superior accuracy, sensitivity, and specificity across all metrics. Conclusion: This study validates the superiority of the ultrasound image-based lightweight deep learning model (ResNet-50) in predicting stroke risk in patients with carotid plaques, making it a preferred clinical diagnostic tool.