DMFNet: a semi-supervised dual-path multiscale fusion network for lesion segmentation in breast ultrasound

Abstract

Accurate lesion segmentation in breast ultrasound images is crucial for early diagnosis and treatment planning, yet remains highly challenging due to inherent imaging characteristics, including low signal-to-noise ratio, speckle noise artifacts, blurred boundaries, and significant inter-lesion appearance variability. Furthermore, the scarcity of expert-annotated medical data severely limits the development of robust deep learning models. To address these challenges, this paper proposes DMFNet, an end-to-end semi-supervised segmentation framework that synergistically integrates the Segment Anything Model (SAM) with a novel dual-path multi-scale fusion architecture. The main contributions include: (1) leveraging a SAM pre-trained Vision Transformer (ViT-B) encoder to extract hierarchical features with global context awareness; (2) designing an ASPP multi-scale context aggregation module at the network bottleneck to accommodate differentiated lesion sizes; (3) cascading channel-spatial dual attention at each decoding stage to focus on lesions and suppress speckle noise; (4) proposing a semi-supervised learning strategy combining confidence threshold pseudo-labeling with weak-strong consistency regularization. Experiments on the internal GD dataset and the external BUSI dataset demonstrate that the proposed method is highly competitive, and ablation experiments further validate its effectiveness.