DABS-Net: A Dual-branch Adaptive Boundary-Structure Network for carotid intima-media complex segmentation

Accurate segmentation of the carotid intima-media complex (IMC) in ultrasound images is essential for carotid intima-media thickness measurement and atherosclerosis assessment. However, IMC segmentation remains challenging because the target is slender, low-contrast, and often affected by speckle noise, blurred interfaces, and plaque-related artifacts, making it difficult to preserve structural continuity and boundary accuracy simultaneously. In this paper, we propose DABS-Net, a lightweight Dual-branch Adaptive Boundary-Structure Network for carotid IMC segmentation. DABS-Net uses an attention-enhanced main branch to capture the continuous IMC region and a gated-convolution-based edge branch to enhance weak lumen-intima and media-adventitia boundary information. An adaptive fusion strategy is further introduced to integrate structural segmentation and boundary refinement features. Experiments on the public CUBS1 and CUBS2 datasets demonstrate the effectiveness of DABS-Net. The proposed method achieves a Dice score of 89.9% on CUBS1 five-fold cross-validation and 88.4% in cross-dataset evaluation, with only 3.58M parameters. Boundary-aware and continuity-related evaluations further show that DABS-Net produces more complete and boundary-consistent segmentation under challenging ultrasound conditions.