UT-MT: A Semi-Supervised Model of Fusion Transformer for 3D Medical Image Segmentation

The training of 3D medical image segmentation model requires a large amount of labeled data, but the availability of labeled data is difficult, and the scarcity of labeled data makes the prediction quality of unlabeled data cannot be effectively guaranteed. To solve the above problems, the 3D medical image segmentation model UT-MT proposed in this paper combines ViT and CNN, and consists of a student model and a teacher model, and the student model learns the teacher model by minimizing the segmentation loss and consistency loss. By combining the feature learning advantages of CNN and ViT, this method enables the model to enhance the learning ability of taking into account both local and global aspects in feature extraction, and further improves the model accuracy and performance. The evaluation was performed on a public left atrial benchmark dataset, and the results at 10% of the labeled images show that the proposed method improves the Dice coefficient by 6.69% over the fully supervised method and has better segmentation of the boundaries, while our method outperforms five advanced semi-supervised segmentation methods.