Swin Unet3D: a three-dimensional medical image segmentation network combining vision transformer and convolution

Abstract

Abstract Background Semantic segmentation of brain tumors plays a critical role in clinical treatment, especially for three-dimensional (3D) magnetic resonance imaging, which is often used in clinical practice. Automatic segmentation of the 3D structure of brain tumors can quickly help physicians understand the properties of tumors, such as the shape and size, thus improving the efficiency of preoperative planning and the odds of successful surgery. In past decades, 3D convolutional neural networks (CNNs) have dominated automatic segmentation methods for 3D medical images, and these network structures have achieved good results. However, to reduce the number of neural network parameters, practitioners ensure that the size of convolutional kernels in 3D convolutional operations generally does not exceed $$7 \times 7 \times 7$$ 7 × 7 × 7 , which also leads to CNNs showing limitations in learning long-distance dependent information. Vision Transformer (ViT) is very good at learning long-distance dependent information in images, but it suffers from the problems of many parameters. What’s worse, the ViT cannot learn local dependency information in the previous layers under the condition of insufficient data. However, in the image segmentation task, being able to learn this local dependency information in the previous layers makes a big impact on the performance of the model. Methods This paper proposes the Swin Unet3D model, which represents voxel segmentation on medical images as a sequence-to-sequence prediction. The feature extraction sub-module in the model is designed as a parallel structure of Convolution and ViT so that all layers of the model are able to adequately learn both global and local dependency information in the image. Results On the validation dataset of Brats2021, our proposed model achieves dice coefficients of 0.840, 0.874, and 0.911 on the ET channel, TC channel, and WT channel, respectively. On the validation dataset of Brats2018, our model achieves dice coefficients of 0.716, 0.761, and 0.874 on the corresponding channels, respectively. Conclusion We propose a new segmentation model that combines the advantages of Vision Transformer and Convolution and achieves a better balance between the number of model parameters and segmentation accuracy. The code can be found at https://github.com/1152545264/SwinUnet3D .