Mitigating over-smoothing in Mamba via high-frequency-guided feature optimization for medical image segmentation

Abstract State Space Models (SSMs) have demonstrated remarkable potential in medical image segmentation owing to their efficient global context modeling capabilities. However, the sequential scanning mechanism of SSMs tends to induce feature over-smoothing, attenuating high-frequency lesion details and blurring irregular boundaries. In this paper, we propose HFG-Mamba, a Vision Mamba network architecture driven by high-frequency-guided feature optimization. With marginal computational and parameter overhead, the proposed architecture achieves a synergy between high-frequency prior extraction and boundarysemantic interaction. Specifically, a high-frequency prior guidance branch built upon the Discrete Wavelet Transform provides the backbone with high-frequency spatial priors. The Spatial Recalibration Fusion Block leverages these priors to counteract over-smoothing by recalibrating backbone features. The decoder further incorporates a Boundary-Semantic Interaction Module to restore boundary details degraded during upsampling. Moreover, a structure-preserving deep supervision strategy is introduced to address the optimization difficulties arising from gradient vanishing and boundary sparsity. Experiments on the ISIC17, ISIC18, and Kvasir-SEG datasets indicate that HFG-Mamba exhibits competitive performance, effectively delineating irregular boundaries and suppressing background artifacts.