Dynamic adaptation and boundary-aware learning for semi-supervised medical image segmentation.
TL;DR
DABAL is proposed, a semi-supervised framework designed to improve both supervision reliability and contour localization and introduces a Dynamic-static Domain Adaptive Adapter (DDAA) into the Segment Anything Model (SAM) encoder to preserve stable structural priors while providing input-dependent feature compensation for knowledge distillation.
Abstract
Limited annotated data remains a major challenge in medical image segmentation. Existing semi-supervised methods are often sensitive to appearance mismatch between labeled and unlabeled images, especially near ambiguous boundaries. To address this problem, we propose Dynamic Adaptation and Boundary-Aware Learning for Semi-supervised Medical Image Segmentation (DABAL), a semi-supervised framework designed to improve both supervision reliability and contour localization. Specifically, we introduce a Dynamic-static Domain Adaptive Adapter (DDAA) into the Segment Anything Model (SAM) encoder to preserve stable structural priors while providing input-dependent feature compensation for knowledge distillation. This adaptation alleviates the effect of appearance variation between labeled and unlabeled samples and makes feature transfer more stable. We further develop Boundary-Aware Supervision (BAS), which imposes explicit constraints on uncertain boundary regions and helps reduce error accumulation during semi-supervised training. Experiments show that DABAL achieves the best performance on ACDC and the best overall performance across the five colonoscopy datasets. Compared with KnowSAM, on ACDC with 10% labeled data, the proposed method improves Dice by 0.95% and IoU by 1.32%, while reducing HD95 by 0.23 and ASD by 0.14. Across the five colonoscopy benchmarks, it improves mean Dice by 3.28% and mean IoU by 3.60%, while reducing mean HD95 by 0.19.