HAF-Net: A Fully Convolutional Segmentation Network Based on Hybrid Attention for Skin Lesion Segmentation

Abstract

Skin cancer has become one of the leading causes of life threatening, and accurate segmentation of lesion regions from dermoscopic images is an effective aid to diagnosis. The lack of obvious lesion features and the large variations in the shape of lesion areas lead to the inability of many convolutional neural networks (CNNs) to accurately identify lesion pixels. In this paper, we propose a fully convolutional segmentation network based on hybrid attention (HAF-Net) for skin lesion segmentation. The model integrates the residual-dense block (RDB) in the encoding stage to improve the efficiency of feature extraction. Furthermore, channel attention block (CAB) and position attention block (PAB) are embedded at the end of the encoding stage, and the semantic relevance of the same class of features is integrated by hybrid attention to highlight the feature representation. In addition, we use channel separable convolution to reduce the number of parameters of the model. Test results on the publicly available skin lesion datasets ISIC 2016 and ISIC 2017 illustrate the superiority of the proposed model.