WSSM : A Weakly Supervised Oral Mucosal Disease Segmentation Model Based on Multi‐Task Collaboration

BACKGROUND: Traditional oral mucosal disease (OMD) diagnosis relies heavily on clinicians' experience and visual assessment, suffering from high subjectivity and low efficiency. OMD images also have insufficient supervision information and fuzzy lesion boundaries, failing to meet mobile medicine's high accuracy requirements. METHODS: To solve these issues, we proposed a weakly supervised OMD segmentation model with multi-task collaboration (WSSM). Using Mamba as the backbone, WSSM realizes efficient lesion segmentation via classification-segmentation dual-branch collaboration. The classification branch captures multi-directional, multi-scale features via a dedicated network, and its pseudo-label module fuses class activation maps, box annotations, and predictive annotations for deeper supervision. The segmentation branch adopts a symmetric network to extract overall lesion features, with a boundary adaptive module enhancing fuzzy boundary representation to improve accuracy. RESULTS: Experiments on the OMD dataset demonstrated that WSSM outperformed existing weakly supervised methods significantly, with its Dice index increasing by 6.06% compared to WSSL. CONCLUSIONS: Our model, with Mamba as the backbone (balancing local texture feature extraction and long-range semantic dependency modeling of OMD lesions), enables deeper supervision via dual-branch collaboration, significantly improving boundary segmentation accuracy in scenarios with insufficient OMD supervision and unclear boundaries. PAPER CODE: https://github.com/XJ156/WSSM3.git.