Cascaded Training Pipeline for 3D Brain Tumor Segmentation

We apply a cascaded training pipeline for the 3DU-Net to segment each brain tumor sub-region separately and chronologically. Firstly, the volumetric data of four modalities are used to segment the whole tumor in the first round of training. Then, our model combines the whole tumor segmentation with the mpMRI images to segment the tumor core. Finally, the network uses whole tumor and tumor core segmentations to predict enhancing tumor regions. Unlike the standard 3D U-Net, we use Group Normalization and Randomized Leaky Rectified Linear Unit in the encoding and decoding blocks. We achieved dice scores on the validation set of 88.84, 81.97, and 75.02 for whole tumor, tumor core, and enhancing tumor, respectively.