Brain tumor segmentation with self-ensembled, deeply-supervised 3D U-net neural networks: a BraTS 2020 challenge solution

TL;DR

This paper presents a novel ensemble-based deep learning approach using self-ensembled, deeply-supervised 3D U-net neural networks for brain tumor segmentation, achieving top-tier results in the BraTS 2020 challenge.

Contribution

The authors introduce a self-ensembled, deeply-supervised 3D U-net framework with model merging strategies, advancing brain tumor segmentation accuracy over prior methods.

Findings

Achieved Dice scores of 0.81, 0.91, 0.85 on validation set.

Ranked among the top ten teams in BraTS 2020.

Open-sourced the segmentation solution for community use.

Abstract

Brain tumor segmentation is a critical task for patient's disease management. In order to automate and standardize this task, we trained multiple U-net like neural networks, mainly with deep supervision and stochastic weight averaging, on the Multimodal Brain Tumor Segmentation Challenge (BraTS) 2020 training dataset. Two independent ensembles of models from two different training pipelines were trained, and each produced a brain tumor segmentation map. These two labelmaps per patient were then merged, taking into account the performance of each ensemble for specific tumor subregions. Our performance on the online validation dataset with test time augmentation were as follows: Dice of 0.81, 0.91 and 0.85; Hausdorff (95%) of 20.6, 4,3, 5.7 mm for the enhancing tumor, whole tumor and tumor core, respectively. Similarly, our solution achieved a Dice of 0.79, 0.89 and 0.84, as well as Hausdorff (95%) of 20.4, 6.7 and 19.5mm on the final test dataset, ranking us among the top ten teams. More complicated training schemes and neural network architectures were investigated without significant performance gain at the cost of greatly increased training time. Overall, our approach yielded good and balanced performance for each tumor subregion. Our solution is open sourced at https://github.com/lescientifik/open_brats2020.