SMRU-Net: skin disease image segmentation using channel-space separate attention with depthwise separable convolutions

Abstract

Skin disease image segmentation faces two major challenges: the complex and varied lesion morphology and the presence of interfering image backgrounds. To address these difficulties in skin disease image segmentation, we propose a Residual U-Net architecture with Channel-Space Separate Attention based on depthwise separable convolutions. The multi-scale residual U-Net modules in the encoder efficiently capture multi-scale texture information in lesions and backgrounds within a single stage, overcoming the limitations of U-Net in extracting just local features. The introduction of ConvMixer Block for global contextual modeling contributes to suppress complex background interference and enhances the overall understanding of lesion morphology. Additionally, we employ a Channel-Space Separate Attention mechanism with depthwise separable convolutions(CSSA-DSC) for feature fusion, effectively addressing the limited expressiveness issue associated with U-Net’s direct skip-connection concatenation. Experimental results on the PH2, ISIC 2017, and ISIC 2018 datasets demonstrate our method’s strong multi-scale modeling and feature expression capabilities.