Skin lesion image segmentation based on improved U-shaped network

Abstract

Skin lesion segmentation is an important step to help us in skin cancer diagnosis. However, skin lesions have variable shapes, fuzzy boundaries, and various colors, which bring challenges to the accurate segmentation of skin lesions. In this paper, we propose an Enhanced Deep Information Fusion based on residual U-net (ED-ResUnet). Firstly, the coding part is improved into a residual network structure, and more feature information is obtained by stacking residual blocks. Secondly, a Channel shuffling hybrid dilated convolution module is designed to increase the receptive field, so that the network pays more attention to deep feature information and strengthens the transmission of context information. Then, the context fusion module and Polarized Self-Attention module are introduced into the encoder to suppress interference information and retain effective lesion location information. Finally, the experimental results on the ISIC 2017 dermoscopy image dataset show that the Dice Coefficient of the segmentation results of the proposed algorithm reaches 87.43%, which is 8.81% higher than that of the original U-Net network and 5.7% higher than that of the UNet++ network. It can segment melanoma effectively.