Multi-view hierarchical split network for brain tumor segmentation

Abstract The use of computer-aided diagnosis in magnetic resonance images to segment brain tumors is clinically crucial for the treatment and rehabilitation of patients. Deep learning has shown great promise in image-segmentation tasks. The existing model is based on UNet, which relies on 3D convolution to learn a large amount of image information to extract feature information. However, UNet has a large number of network parameters, and using only 3D convolution as a feature extraction module limits the efficiency and effectiveness of the network. In this study, we proposed a multi-view hierarchical split network (MVHS-Net) for brain segmentation. The main contribution of this study is the design of a new lightweight encoder-decoder structure called modified 3D UNet. Regarding this structure, multi-view fusion convolution and multi-view hierarchical split block (MVHS block) are proposed. These modules can efficiently obtain multi-scale and multi-view information, reduce redundant characteristic information, and improve network performance. Numerous experimental results that are achieved via the BraTS 2018 challenge dataset show that the proposed architecture significantly reduces the computational complexity while maintaining the high precision of brain tumor segmentation. The proposed approach achieves dice scores of 89.55%, 85.24% and 80.22%, and Hausdorff Distances (95th percentile) of 2.32, 5.21, 5.71 for the WT, ET, and TC, respectively. At the same time, the parameter amount of our proposed method is only 0.35M. Compared to the state-of-the-art lightweight brain tumor segmentation algorithms, our proposed network has fewer parameters and more accurate segmentation results.