Skin lesion segmentation combining feature refinement and context guide

Abstract

Aiming at the problem of high-precision segmentation of skin lesions, a skin lesion segmentation network combining feature refinement and context guide is proposed. Firstly, a dual-layer feature thinning module is designed to mine the difference information and common information between adjacent feature layers, and generate weight vectors to guide the encoder feature map to gradually refine, so as to enhance the ability of feature expression. Secondly, a dense residual pyramid context guide module is designed at the highest level of the network to expand the network’s receptive field through cascading expansion convolution, and integrate features of different scales in a hierarchical residual connection method to achieve dense aggregation of spatial information, and then combine global and local attention establish a multi-scale and multi-dimensional context prior to guiding the network to pay more attention to the target area and reduce noise interference. Finally, the cross-entropy loss and weighted boundary loss are combined to supervise the shape of the lesion in the model training process to improve the accuracy of boundary prediction.