Efficient 3D depthwise and separable convolutions with dilation for brain tumor segmentation

In this paper, we propose a 3D convolutional neural network targeting at the segmentation of brain tumor. There are different types of brain tumors and our focus is one common type named glioma. The proposed network is efficient and balances the tradeoff between the number of parameters and accuracy of segmentation. It consists of Anisotropic Block, Dilated Parallel Residual Block, and Feature Refinement Module. The Anisotropic Block applies anisotropic convolutional kernels on different branches. In addition, the Dilated Parallel Residual Block incorporates 3D depthwise and separable convolutions to reduce the amount of required parameters dramatically, while multiscale dilated convolutions enlarge the receptive field. The Feature Refinement Module prevents global contextual information loss. Our method is evaluated on the BRATS 2017 dataset. The results show that our method achieved competitive performance among all compared methods, with a reduced number of parameters. The ablation study also proves that each individual block or module is effective.