Multimodal Brain Tumor Image Segmentation Based on 3D CTrans U-Net: Replacing Skip Connection with Channel-Wise Attention Module

Automatic segmentation of brain tumors can be used as a powerful auxiliary means for clinical diagnosis, bringing better treatment. In recent years, medical image segmentation technology has mainly developed into two genres: convolution-based and Transformer-based. However, both of them more focus on the design of feature extraction, while less on the role of the skip-connection in the network structure. Based on the analysis of the contributions and disadvantages of skip-connection, the 3D CTrans U-Net is proposed to apply channel-wise cross fusion Transformer (CCT) and channel-wise cross attention (CCA) module to replace the skip-connection in 3D U-Net, which can better integrate multi-scale encoder features and bridge the semantic gap between encoder and decoder. The proposed method has achieved more accurate segmentation results on the BraTS2019 dataset, which has been improved to a certain extent compared with other state-of-the-art methods.