DSU-Net: Dual-Stage U-Net based on CNN and Transformer for skin lesion segmentation

Abstract

Precise delineation of skin lesions from dermoscopy pictures is crucial for enhancing the quantitative analysis of melanoma. However, this remains a difficult endeavor due to inherent characteristics such as large variability in lesion size, form, and fuzzy boundaries. In recent years, CNNs and Transformers have indicated notable benefits in the area of skin lesion segmentation. Hence, we first propose the DSU-Net segmentation network, which is inspired by the manual segmentation process. Through the coordination mechanism of the two segmentation sub-networks, the simulation of a process occurs where the lesion area is initially coarsely identified and then meticulously delineated. Then, we propose a two-stage balanced loss function to better simulate the manual segmentation process by adaptively controlling the loss weight. Further, we introduce a multi-feature fusion module, which combines various feature extraction modules to extract richer feature information, refine the lesion area, and obtain accurate segmentation boundaries. Finally, we conducted extensive experiments on the ISIC2017, ISIC2018, and PH2 datasets to assess and validate the efficacy of the DSU-Net by comparing it to the most advanced approaches currently available. The codes are available at https://github.com/ZhongLongwei/DSU-Net.