Improved U-Net based on contour attention for efficient segmentation of skin lesion.

In the field of skin disease diagnosis, the segmentation of lesions is a crucial step in auxiliary diagnosis. The U-Net network has achieved many successes in the field of medical segmentation. However, due to the relatively simple encoder and decoder of traditional U-Net, it is easy to be affected by factors such as small sample size, insufficient feature expression, and irregular shape of segmented targets, the good features cannot be extracted, resulting in incorrect segmentation results. In order to solve these problems and improve the performance of lesion segmentation. We designed a new U-Net network by improving the traditional architecture in two ways: 1) We replaced the encoder with a robust classification network to effectively extract feature information. 2) To thoroughly exploit the potential perceptual clues in the multi-scale feature maps generated by the encoder, we designed the Multi-Scale Depth Feature Exploration (MSD) module as the decoder. Additionally, we propose a novel segmentation loss function named Weighted Contour Cross Entropy loss (WCCE), which forces the network pay more attettion to lesion contour information and improve the contour quality of lesion segmentation to achieve better segmentation accuracy. All experiments were conducted on two skin lesion datasets, ISIC2018 and XJUSL. Experimental analysis reveals that the improved U-Net assisted with proposed loss surpasses the traditional U-Net and other typical methods in the performance of all evaluation metrics used in this study. This result reflects that the proposed method can effectively segment the lesion contour and improve the performance of lesion segmentation. [ABSTRACT FROM AUTHOR]