Transclaw U-Net: Claw U-Net With Transformers for Medical Image Segmentation

In medical image analysis, the long-range spatial features are often not accurately obtained by the traditional convolutional neural networks. Hence, we propose a TransClaw U-Net network structure. The transformer part is added after three convolution operations to fuse shallow features extracted by convolution operations for maximally encoding the long-range spatial features between patches. The "Claw" in TransClaw U-net means that we add the bottom upsampling part to retain the deepest feature information for detail segmentation. In addition, the modified three-channel global attention mechanism to blend the outputs of three channels (the encoding part, the bottom upsampling part and the decoding part) to effectively extract image contours. The experimental results on Synapse Multi-organ Segmentation Dataset show that TransClaw U-Net performs better than other networks. The results of ablation experiments prove the effectiveness of the three improved components of the network and influence of input image size and skip-connection numbers on network performance. The source code will be publicly available once the paper is accepted.