Abstract
<title>Abstract</title> <p>Accurate segmentation of breast lesions in ultrasound images is critical for early cancer diagnosis, yet it remains challenging due to speckle noise, acoustic artifacts, and blurred boundaries. In this paper, we propose DMPB-Net, a novel semi-supervised framework designed for precise breast lesion segmentation under limited supervision. The framework utilizes a Mamba-driven encoder to capture long-range global dependencies, effectively overcoming the receptive field limitations of conventional CNNs. To mitigate ultrasound-specific interference, we introduce a Semantic Decoupling (SD) module that separates lesion-related features from task-irrelevant background noise. Furthermore, a Confidence-Ranked Prototype Attention (CR-PA) module is developed to facilitate robust semantic alignment by dynamically filtering high-confidence prototypes. To address boundary ambiguity, a Boundary Iterative Refinement (BIR) module is employed to progressively polish masks through a generative denoising process. Finally, a composite loss function comprising Consistency, Fuzzy Rough Set Boundary (FRSB), and Structural Similarity Contrastive (SSC) losses is introduced to supervise both local details and global topology. Extensive experiments on BUSI and UDIAT datasets demonstrate that DMPB-Net significantly outperforms state-of-the-art methods, achieving a Dice score of 83.20% and 91.51% respectively, showcasing its high clinical potential for computer-aided diagnosis. Code is available at https://github.com/HengfanLi/DMPB-Net2.</p>