3D Automatic Brain Tumor Segmentation Using a Multiscale Input U-Net Network

Quantitative analysis of brain tumors is crucial for surgery planning, follow-up and subsequent radiation treatment of glioma. Finding an automatic and reproducible solution may save time to physicians and contribute to improve overall poor prognosis of glioma patients. In this paper, we present our current BraTS contribution on developing an accurate and robust tumor segmentation algorithm. Our network architecture implements a multiscale input module which has been thought to maximize the extraction of features associated to the multiple image modalities before they are merged in a modified U-Net network avoiding the loss of specific information provided by each modality and improving brain tumor segmentation performance. Our method's current performance on the BraTS 2019 test set is dice scores of 0.775 +/- 0.212, 0.865 +/- 0.133 and 0.789 +/- 0.266 for enhancing tumor, whole tumor and tumor core, respectively with and overall dice of 0.81.