SHARED-WEIGHT SEGMENTATION WITH MASK-CONDITIONED RISK CLASSIFICATION FOR CAROTID ULTRASOUND

Stroke is a leading cause of mortality and long-term disability worldwide, with carotid atherosclerosis serving as a major contributor to ischemic events. Carotid ultrasound is routinely used for plaque morphology assessment. However, automated plaque risk grading remains challenging due to small plaque size, low image contrast, and indistinct plaque-lumen boundaries. Moreover, longitudinal and transverse views provide complementary anatomical information that requires coherent cross-view reasoning, which is insufficiently leveraged in the existing automated pipelines. To address these challenges, we propose a two-stage, two-view framework that integrates semi-supervised plaque and vessel segmentation with mask-informed risk classification. The model is trained using 200 labeled cases and 800 unlabeled cases. On the challenge evaluation, our method achieves an F1-Score of 65.50 and a Seg-Score of 62.99, outperforming the official baseline (F1 33.96, Seg-Score 54.71) by 31.54 and 8.28, respectively. These results demonstrate the effectiveness of integrating multi-view anatomical priors with semi-supervised learning for the precise grading of carotid atherosclerotic risk. Code is publicly available at https://github.com/06Yujie/CSV2026-Challenge.