LightMamba-UNet: Lightweight Mamba with U-Net for Efficient Skin Lesion Segmentation

Wanzhen Hou, Shiwei Zhou, Haifeng Zhao

Abstract: Recently, advancements in Vision Mamba Models (VMMs) have achieved notable success in dense prediction. However, existing Vision-Mamba-like approaches often improve performance by adopting more complex modules, making them poor for medical applications where computational resources are constrained. Meanwhile, focusing on segmentation tasks, these methods prefer to redesign the internal architecture of the encoder and decoder, potentially overlooking the thorough exploration of skip connections, which in turn affects the accuracy of segmentation. To minimize the parameters of the model derived from Vision Mamba, we introduce a novel Lightweight Mamba with U-Net (LightMamba-UNet) for processing the deep features in parallel. Furthermore, to address the demand for more accurate segmentation, we introduce a full-scale lightweight skip connection to bridge the semantic gap between the encoder and decoder. Extensive experiments demonstrate that our model outperforms the state-of-the-art (SOTA) on the public ISIC17, ISIC18, and \(\text {PH}^2\) datasets, concurrently the proposed parameter efficient framework can reduce the total model size by 36.73\(\%\).