EG-TransUNet: Enhanced and Guided U-Net with Transformer for Biomedical Image Segmentation

Abstract Although methods based on convolutional neural network have improved the performance of biomedical image segmentation, to meet the precision requirements of medical imaging segmentation task, the medical image segmentation method based on the deep learning still need to solve the following problems: 1) It is difficult to extract the discriminative feature of the lesion region in medical images during the encoding caused by variable sizes and shapes; 2) It is difficult to fuse the spatial and semantic information of the lesion region effectively during the decoding caused by redundant information and semantic gap. In this paper, we propose a novel U-Net variant architecture called EG-TransUNet, which is able to improve the feature discrimination at the level of spatial detail and semantic location using progressive enhancement module (PEM) and channel spatial attention (CSA) based on self-attention and effectively fuse the spatial and semantic information using semantic guidance attention (SGA). The proposed EG-TransUNet allows to capture object variabilities and provides improved results on different biomedical datasets. Extensive experiments on EG-TransUNet demonstrate that the method advances the performance on five publicly available segmentation datasets, and also, is more generalizable as compared to state-of-the-art methods.