Ligs-Mamba: a synergistic Mamba network for video breast Cooper’s ligament segmentation

Accurate segmentation of breast Cooper’s (BC) ligaments in endoscopic mastectomy videos is difficult because of their low contrast, indistinct boundaries, and interference from surrounding tissues. Reliable delineation supports anatomical identification during surgery. Mamba has shown promise in medical image analysis. Compared with Transformer self-attention, its selective scanning mechanism adaptively emphasizes informative features in long sequences. We propose Ligs-Mamba, a Mamba-based framework for video BC ligament segmentation. It comprises an anatomy-aware global Mamba (AGM) module and an uncertainty-focused local Mamba (ULM) module. AGM uses coarse ligament predictions to direct global feature scanning toward anatomically relevant regions, whereas ULM uses uncertainty priors to focus local scanning on ambiguous boundaries. Together, these modules strengthen the representation of low-contrast ligaments while reducing background interference. We also collect and annotate the video-based BC ligament segmentation (VLS) dataset, which contains 90 videos and 6403 frames. To our knowledge, VLS is the first annotated dataset for video BC ligament segmentation, and Ligs-Mamba is the first framework designed for this task. Experiments on VLS show that Ligs-Mamba outperforms state-of-the-art methods in quantitative and qualitative evaluations, producing more accurate ligament masks under challenging conditions.