Skin Lesion Segmentation via Improved U-Net with Spatial Group-wise Enhancement and Multi-scale Parallel Feature Fusion

Abstract

Accurate segmentation of skin lesions is a prerequisite for automated dermatological diagnosis. While the U-Net architecture is widely used for medical image segmentation, its performance is often limited by background noise interference and the loss of multi-scale context in deep layers. This paper proposes an improved U-Net-based model tailored for skin lesion segmentation. We integrate a lightweight Spatial Group-wise Enhancement (SGE) attention mechanism into the encoder to suppress non-pathological textures and noise. Furthermore, a Multi-scale Parallel Feature Fusion (MPFF) module is introduced at the deep stages of the network to aggregate multi-scale semantic information and preserve high-frequency boundary details. Experimental results on the ISIC benchmarks show that our proposed model significantly outperforms the baseline U-Net in terms of Dice Score and Intersection over Union (IoU), providing a robust tool for clinical skin lesion analysis.