MRI Brain Tumor Segmentation Using Deep Encoder-Decoder Convolutional Neural Networks

In this study, we focus on Task 1 of the 2021 Multimodal Brain Tumor Segmentation (BraTS) challenge. We present a modified U-net model aimed at improving the segmentation of glioblastomas, reducing the computation timewithout compromising detection sensitivity. Our automated approach takes multimodal MR images as input, generates a bounding box of the brain volume, and combines the model predictions at the 2D slice level into a full 3D segmentation that is written into a NIfTI file. On the official 2021 BraTS test set of 570 cases, the model obtained median Dice scores of 0.80, 0.87, and 0.87, as well as median 95% Hausdorff distances of 2.45, 4.64, and 6.40 for the enhancing tumor, tumor core, and whole tumor regions, respectively.