Ensemble Outperforms Single Models in Brain Tumor Segmentation.

Brain tumor segmentation remains an open and popular challenge, for which countless medical image segmentation models have been proposed. Based on the platform that BraTS challenge 2021 provided for researchers, we implemented a battery of cutting-edge deep neural networks, such as nnU-Net, UNet++, CoTr, HRNet, and Swin-Unet to directly compare performances amongst distinct models. To improve segmentation accuracy, we first tried several modification techniques (e.g., data augmentation, region-based training, batch-dice loss function, etc.). Next, the outputs from the five best models were averaged using a final ensemble model, of which four models in the committee were organized in different architectures. As a result, the strengths of every single model were amplified by the aggregation. Our model took one of the best performing places in the Brain Tumor Segmentation (BraTS) 2021 competition amongst over 1200 excellent researchers from all over the world, which achieved Dice score of 0.9256, 0.8774, 0.8576 and Hausdor Distances (95%) of 4.36, 14.80, 14.49 for whole tumor, tumor core, and enhancing tumor respectively.