Skin lesion segmentation is essential for early disease detection and treatment planning in computer-aided diagnostic systems. However, U-Net faces challenges in handling long-distance dependencies and fully utilizing semantic information. Additionally, feature redundancy in channels and asymmetric supervised learning can lead to irrelevant features, resulting in suboptimal segmentation accuracy. To tackle these challenges, this paper presents a dermatological segmentation method that improves inter-channel and spatial features. The method introduces a compression excitation module and a channel mixing network, boosting both feature extraction capabilities and channel information exchange. Furthermore, the integration of a cross-region attention mechanism enhances the modeling of long-distance dependencies and spatial feature perception. The proposed approach also integrates a feature distillation loss function, which facilitates a balanced supervision mechanism between the encoder and decoder. This effectively minimizes redundant information within the U-Net architecture. Experiments conducted on the publicly available skin lesion datasets ISIC2016, ISIC2017, and ISIC2018 demonstrate that the proposed approach attains substantial performance enhancements in skin lesion image segmentation, showcasing its strong competitiveness.