Accurate skin lesion segmentation remains a significant challenge for lightweight models, as they often struggle to effectively balance global context understanding with stable local structural representation. Shallow skip connections can introduce redundant background noise into the decoder, while lightweight Mamba-based backbones have limited ability to preserve two-dimensional local continuity during sequential scanning. Additionally, direct enhancement in the frequency domain may amplify high-frequency components that are irrelevant to lesion discrimination, thereby compromising representation stability. To address these issues, this paper proposes LSF-Mamba, a lightweight skin lesion segmentation framework that integrates spatial-domain local compensation and frequency-domain residual supplementation. Specifically, a Local Context Recalibration (LCR) module is designed to reinforce shallow structural features through local neighborhood aggregation and channel-wise recalibration. In parallel, a Frequency Residual Enhancement (FRE) module is introduced to selectively incorporate structure-aware multi-band information using locally aligned wavelet decomposition and gated residual injection. Extensive experiments on the ISIC2017, ISIC2018, and PH2 datasets demonstrate that the proposed method achieves competitive performance with only 0.0491 M parameters and 0.028 G FLOPs, while producing more complete lesion segmentation and more accurate boundary delineation.
Zhang et al. (Mon,) studied this question.