A skin lesion segmentation network with edge and body fusion

Purpose: In this paper, a novel local cross-attention Unet(LCAUnet) is proposed to enhance the completeness of representation with the fusion of edge and body features, which are often paid little attention in traditional methods. Method: First, two separate branches are set for edge and body segmentation with convolutional neural networks(CNNs) and Transformer based architectures, respectively. Then, the local cross-attention feature fusion(LCAF) module is utilized to merge feature maps of the edge and body of the same level via local cross-attention operation in the encoder stage, and the edge-body interactions can be captured hierarchically. Furthermore, the prior guided multi-scale knowledge fusion(PGMF) module is embedded for feature integration with prior guided multi-scale adaption. Result: Comprehensive experiments on publicly available datasets ISIC 2017, ISIC 2018, and PH2 demonstrate that LCAUnet outperforms most state-of-the-art methods in metrics such as 1.31% improvement in Dice. The ablation studies also verify the effectiveness of the proposed fusion techniques.