NnUNet with Region-based Training and Loss Ensembles for Brain Tumor Segmentation

Abstract

Brain tumor segmentation in multi-model MRI scans is a long-term and challenging task. Motivated by the winner solution in BraTS 2020 [7], we incorporate region-based training, a more aggressive data augmentation, and loss ensembles to build the widely used nnUNet model. Specifically, we train ten cross-validation models based on two compound loss functions and select the five best models for ensembles. On the final testing set, our method achieves average Dice scores of 0.8760, 0.8843, and 0.9300 and 95% Hausdorff Distance values of 12.3, 15.3, and 4.75 for enhancing tumor, tumor core, and whole tumor respectively.