FuSS-Net: Multi-level fusion of frequency priors and semantic state space dynamics for synergistic lesion segmentation and classification

Joint lesion segmentation and classification are important for computer-aided diagnosis, yet existing multi-task frameworks often struggle with ambiguous lesion boundaries, weak or heterogeneous textures, and cross-task interference between dense prediction and image-level recognition. To address these issues, we propose FuSS-Net, a multi-level fusion framework that integrates frequency-domain priors, semantic adaptive state space propagation, and non-additive feature fusion for synergistic lesion analysis. Specifically, a fuzzy rule-based adaptive frequency filtering module is designed to adaptively inject frequency-domain priors into shallow spatial representations through frequency-domain modulation, thereby enhancing boundary and texture sensitive features for both segmentation and classification. A semantic adaptive fuzzy-gated state space module is further introduced to fuse category priors with token propagation, thereby enhancing semantic-aware feature interaction and improving cross-task collaboration. In addition, a texture-enhanced Choquet fusion module performs non-additive fusion of shallow texture and deep semantic features, enabling more effective modeling of complementary information for robust classification. Extensive experiments on the oral ulcer dataset, ISIC 2017, and BUSI show that FuSS-Net achieves strong joint-task performance, including 90.87% segmentation IoU and 98.44% classification accuracy on the oral ulcer dataset, while also obtaining favorable zero-shot external results on Autooral with 88.09% classification accuracy and 90.34% F1-score. These results demonstrate that the proposed framework provides an effective and interpretable solution for joint lesion segmentation and classification across heterogeneous medical imaging scenarios.