Automatic Skin Lesion Segmentation Based on Gated Channel Attention and Boundary-Guide Transformer

Automatic skin lesion segmentation from dermoscopy images is an essential step in the computer-aided skin diagnosis system. However, it is still a challenging task due to the large variation of melanoma and ambiguous boundaries of skin lesions. In this paper, we propose a skin lesion segmentation network based on gated channel attention and transformer, called the Boundary-guide Transformer Network(BGT-Net). Firstly, we modify ResNet-50 model and introduce the Gated Channel Transformation(GCT) into residual structure to identify feature-rich channels. In addition, we design and integrate the boundary guide module(BGM) into transformer to enhance ambiguous boundary features of skin lesions while effectively modeling global long-range dependencies. Finally, the proposed BGT-Net is evaluated on two published standard datasets, ISIC2016 and ISIC2018. Extensive experimental results demonstrate that the proposed network achieves better segmentation performance over other state-of-the-art methods.