Synergistic perception: Fusing expert knowledge and foundation models for semi-supervised mammogram segmentation

Mammography is essential for the early detection of breast cancer, but accurately segmenting complex tissue structures across varying scales remains challenging due to data scarcity and inherent structural variability. We introduce the Synergistic Perception Framework (SPF), a novel approach that integrates specialized components operating at different scales to enhance segmentation performance. The SPF consists of three key components: (1) Expert Unit Models (EUMs) that capture fine-grained, class-specific details; (2) a Hierarchical Feature Fusion Network (HFF-Net) that integrates deep contextual information with localized features through a category-adaptive feature decoupling decoder; and (3) a progressive pseudo-label refinement strategy that leverages unlabeled data. This process uses consistency regularization for initial pseudo-label generation followed by targeted fine-tuning of the Segment Anything Model (SAM) to produce high-quality segmentation targets. Experimental results demonstrate that SPF outperforms existing methods on the segmentation of 11 anatomical structures across multiple test sets, improving the average Dice score by 13.27 percentage points on CSAW-S and 10.1 percentage points on INbreast compared to state-of-the-art (SOTA) methods. The framework particularly excels in segmenting small and complex structures, validating the effectiveness of our multi-scale approach. The code will be made publicly available upon acceptance.