Abdominal Multi-organ Segmentation using CNN and Transformer

Rui Xin, Lisheng Wang

Keywords: Medical segmentation, Pseudo label, Semi-supervision learning

Abstract: In this paper, we combine the advantages of convolution local correlation and translation invariance in CNN with Transformer’s ability to effectively capture long-term dependencies between pixels to produce high quality pseudo labels. In order to segment images efficiently and quickly, we select nnU-Net as the final segmentation network and use pseudo labels, unlabeled data and labeled data together to train the network, and then we use Generic U-Net, the backbone network of nnU-Net, as final prediction network. Experimental results show that the proposed method achieves excellent semi-supervised segmentation performance in FLARE2022 Challenge.