<p>Skin lesion segmentation aims to accurately identify and segment the lesion areas in medical images. Many traditional lesion segmentation networks have low efficiency in extracting feature information. Moreover, due to the issue of feature information loss during continuous downsampling, directly cascading the encoder features with the network decoder end is not an efficient strategy. To better address these issues, This study presents an innovative Transformer-based U-shape network architecture for accurately segmenting skin lesions, named ESICU-Net. The network adopts Transformer as the encoder, replacing the traditional encoder’s simple stacked convolutional blocks design. This allows for the effective extraction of multi-scale features with rich semantic information. Furthermore, at the network skip-connection points, a spatial feature enhancement (SFE) module is designed. It strengthens the semantic feature information at different levels of the encoder through non-local operations while suppressing irrelevant background noise. The paper also employs a multi-scale channel feature interaction (MCFI) module to comprehensively integrate multi-stage channel feature, achieving the optimized representation of top-level abstract features for bottom-level detailed features. By enhancing and integrating the multi-scale features extracted by the encoder, it efficiently minimizes the semantic feature difference between the network encoder-end and decoder-end, providing rich semantic feature information to the decoding end. In this paper, extensive experimental comparisons were conducted on three skin lesion datasets: ISIC-2018, ISIC-2016, and PH2. Through comparison, ESICU-Net achieved Dice scores of 91.50%, 92.95%, and 91.95% on these three datasets, outperforming other cutting-edge algorithms.</p>