Multimodal Brain Tumor Segmentation Using Encoder-Decoder with Hierarchical Separable Convolution

To address automatic segmentation of brain tumor from multi-modal MRI volumes, a light-weight encoder-decoder network is presented. Exploring effective way to trade off the range of spatial contexts and computational efficiency is crucial to address challenges of 3D segmentation. To this end, we introduce hierarchical separable convolution (HSC), an integration of view- and group-wise separable convolution, which can simultaneously encode multi-scale context in 3D and reduce memory overhead without sacrificing accuracy. Specifically, typical 3D convolution is replaced with complementary 2D convolutions at multiple scales and thus multiple fields-of-view, which results in a light-weight but stronger model. Moreover, thanks to the decomposed convolutions, we ensemble 3D segmentations with different focal views to further improve segmentation accuracy. Experiments on the BRATS 2017 benchmark showed that our method achieved state-of-the-art performance in Dice, i.e., 0.901, 0.809 and 0.762 for the whole tumor, tumor core and enhancing tumor core, respectively.