BRATS2021: Exploring Each Sequence in Multi-modal Input for Baseline U-net Performance

Polina Druzhinina, Ekaterina Kondrateva, Arseny Bozhenko, Vyacheslav Yarkin, Maxim Sharaev, Anvar Kurmukov

Abstract: Since 2012 the BraTS competition has become a benchmark for brain MRI segmentation. The top-ranked solutions from the competition leaderboard of past years are primarily heavy and sophisticated ensembles of deep neural networks. The complexity of the proposed solutions can restrict their clinical use due to the long execution time and complicate the model transfer to the other datasets, especially with the lack of some MRI sequences in multimodal input. The current paper provides a baseline segmentation accuracy for each separate MRI modality and all four sequences (T1, T1c, T2, and FLAIR) on conventional 3D U-net architecture. We explore the predictive ability of each modality to segment enhancing core, tumor core, and whole tumor. We then compare the baseline performance with BraTS 2019–2020 state-of-the-art solutions. Finally, we share the code and trained weights to facilitate further research on model transfer to different domains and use in other applications.