A multi-plane 2D medical image segmentation method combined with transformers

Segmentation on medical image is a necessary prerequisite for disease diagnosis and treatment planning. In various medical image segmentation tasks, U-Net based on convolutional neural network (CNN) has achieved tremendous success due to its ability to learn image details and deep high-dimensional features. However, the inherent limitations of convolution operations limit their performance in modeling explicit long-term relationships. Especially in the brain MRI glioma segmentation task, traditional convolutional neural networks show weakness in learning global semantic information due to the extreme intrinsic heterogeneity of tumors in terms of appearance, shape, and histology. Therefore, this research proposes a 2D U-Net combined with transformer for brain tumor segmentation, and also designs a fusion module to better fuse the high-resolution detail features from the encoder with the high-level semantic features in the decoding process. The respective segmentation models are trained on the three orthogonal planes, and a tumor core-assisted network is separately trained. We validate proposed method on the BraTS20 brain glioma MRI dataset, and the experimental results demonstrate that our method outperforms state-of-the-art 2D methods. Compared with 3D methods, our model can accomplish accurate tumor segmentation while saving computational resources.