Automatic Brain Tumor Segmentation with a Bridge-Unet Deeply Supervised Enhanced with Downsampling Pooling Combination, Atrous Spatial Pyramid Pooling, Squeeze-and-Excitation and EvoNorm

Abstract

AbstractSegmentation of brain tumors is a critical task for patient disease management. Since this task is time-consuming and subject to inter-expert delineation variation, automatic methods are of significant interest. The Multimodal Brain Tumor Segmentation Challenge (BraTS) has been in place for about a decade and provides a common platform to compare different automatic segmentation algorithms based on multiparametric magnetic resonance imaging (mpMRI) of gliomas. This year the challenge has taken a big step forward by multiplying the total data by approximately 3. We address the image segmentation challenge by developing a network based on a Bridge-Unet and improved with a concatenation of max and average pooling for downsampling, Squeeze-and-Excitation (SE) block, Atrous Spatial Pyramid Pooling (ASSP), and EvoNorm-S0. Our model was trained using the 1251 training cases from the BraTS 2021 challenge and achieved an average Dice similarity coefficient (DSC) of 0.92457, 0.87811 and 0.84094, as well as a 95% Hausdorff distance (HD) of 4.19442, 7.55256 and 14.13390 mm for the whole tumor, tumor core, and enhanced tumor, respectively on the online validation platform composed of 219 cases. Similarly, our solution achieved a DSC of 0.92548, 0.87628 and 0.87122, as well as HD95 of 4.30711, 17.84987 and 12.23361 mm on the test dataset composed of 530 cases. Overall, our approach yielded well balanced performance for each tumor subregion.KeywordsDeep-learningBrain tumorSegmentation