GSUNet: A Brain Tumor Segmentation Method Based on 3D Ghost Shuffle U-Net

Research on MRI-based brain tumor segmentation methods has clinical significance and application value. The existing 3D brain tumor segmentation methods can make full use of the three-dimensional spatial information of MRI, but there are problems with large parameters and calculations. In view of the above problems, a brain tumor segmentation method based on 3D Ghost Shuffle U-Net(GSUNet) is proposed. In this paper, the 3D Ghost Module(3D GM) is utilized as the basic feature extractor of the network, which fundamentally solves the problem of the high complexity of the existing 3D brain tumor segmentation models. At the same time, the Ghost Shuffle Module(GSM) is designed, and GSM with stride 2 is utilized to realize down-sampling, optimize the feature extraction, and strengthen the information communication in the channel dimension. The GSM with stride 1 and the designed Dense Ghost Module(DGM) work together as a decoder to improve the representation ability of the network at a lower cost. Experimental results show that GSUNet can achieve segmentation performance comparable to mainstream brain tumor segmentation methods with extremely low model complexity.