Abstract
Accurate classification of carotid plaques is critical to assessing the risk of cardiovascular disease. However, this task remains challenging due to several factors: temporal discontinuity caused by probe motion, the small size of plaques combined with interference from surrounding tissue, and the limited availability of annotated data, which often leads to overfitting in deep learning models. To address these challenges, this study introduces a structured self-distillation framework, named DualDistill, designed to improve classification accuracy and generalization performance in analyzing ultrasound videos of carotid plaques. DualDistill incorporates two novel strategies to address the identified challenges. First, an intra-frame relationship-guided strategy is proposed to capture long-term temporal dependencies, effectively addressing temporal discontinuity. Second, a spatial-temporal attention-guided strategy is developed to reduce the impact of irrelevant features and noise by emphasizing relevant regions within both spatial and temporal dimensions. These strategies jointly act as supervisory signals within the self-distillation process, guiding the student layers to better align with the critical features identified by the teacher layers. Besides, the self-distillation process acts as an implicit regularization mechanism, which decreases overfitting in limited datasets. DualDistill is designed as a plug-and-play framework, enabling seamless integration with various existing models. Extensive experiments were conducted on 317 carotid plaque ultrasound videos collected from a collaborating hospital. The proposed framework demonstrated its versatility and effectiveness. It achieved consistent improvements in classification accuracy across 13 representative models. Specifically, the average accuracy improvement is 2.97%, with the maximum improvement reaching 4.74% on 3D ResNet50. These results highlight the robustness and generalizability of DualDistill. It shows strong potential for reliable cardiovascular risk assessment through automated carotid plaque classification.