TDM-Net: Topology-Enhanced Dynamic Memory Network for Skin Lesion Segmentation

Skin lesion segmentation poses a challenging problem in medical image analysis, with segmentation performance being highly sensitive to blurred boundaries, noise interference, and morphological diversity. To address these challenges, this paper proposes a Topology-Enhanced Dynamic Memory Network (TDM-Net). This model incorporates a topology-aware feature extraction module that integrates Alpha complex construction with persistent homology analysis to explicitly model the connectivity and structural information of lesion regions at the feature level. Additionally, a dynamic feature modulation module is designed to adaptively fuse local window attention with global self-attention, while employing Mahalanobis distance for dynamic calibration of correlations among multi-scale features. Furthermore, a hierarchical memory distillation mechanism is introduced to enable cross-scale feature reuse and historical information aggregation, thereby enhancing feature expression stability and generalization capability in complex noisy environments. The method undergoes systematic validation on multiple public datasets including ISIC 2016, ISIC 2017, ISIC 2018, and PH2. Experimental results demonstrate that TDM-Net achieves 91.76% DSC and 4.35 ASSD on ISIC 2017, yielding a 2.34% DSC improvement and a 15% ASSD reduction over MSDTCN-Net. On ISIC 2018, it reaches 92.13% DSC, outperforming DuaSkinSeg by 2.43% DSC, demonstrating accurate and robust lesion segmentation performance.