MGMFormer: Multi-Scale Attentional Medical Image Segmentation Network for Semantic Feature Enhancement

Multi-scale feature extraction is important for the accurate segmentation of different lesion areas. In order to solve the problem of false cut and missing cut in practical applications due to the difficulty of extracting semantic feature information from existing technologies, we proposed a multi-scale attention network framework based on semantic feature enhancement, MGMFormer. Taking advantage of multi-scale feature extraction and attention mechanism to enhance semantic features, the encoder and decoder are composed of joint learning, multi-scale arbitrary sampling, and global adaptive calibration modules. It makes the encoder more focused on the fine structure, so as to effectively deal with the problem of reduced accuracy caused by modal heterogeneity. At the same time, it solves the problem of lack of feature expression ability when the decoder deals with complex texture information. We evaluated the segmentation performance of MGMFormer on eight different datasets, BraTS, Sypanse, ACDC, ISIC, Kvasir-SEG, CAMUS, CHNCXR, and Glas, and in particular, it outperformed most existing algorithms.