A Memory-efficient Deep Framework for Multi-Modal MRI-based Brain Tumor Segmentation

Automatic Brain Tumor Segmentation (BraTS) from MRI plays a key role in diagnosing and treating brain tumors. Although 3D U-Nets achieve state-of-the-art results in BraTS, their clinical use is limited due to requiring high-end GPU with high memory. To address the limitation, we utilize several techniques for customizing a memory-efficient yet ac-curate deep framework based on 2D U-nets. In the framework, the simultaneous multi-label tumor segmentation is decomposed into fusion of sequential single-label (binary) segmentation tasks. In addition to reducing the memory consumption, it may also improve the segmentation accuracy since each U-net focuses on a sub-task, simpler than whole BraTS segmentation task. Extensive data augmentations on multi-modal MRI and the batch dice-loss function are also employed to further increase the generalization accuracy. Experiments on BraTS 2020 demonstrate that our framework almost achieves state-of-the-art results. Dice scores of 0.905, 0.903, and 0.822 for whole tumor, tumor core, and enhancing tumor are accomplished on the testing set. Moreover, our customized framework is executable on budget-GPUs with minimum requirement of only 2G RAM. Clinical relevance— We develop a memory-efficient deep Brain tumor segmentation tool that significantly reduces the hardware requirement of tumor segmentation while maintaining comparable accuracy and time. These advantages make our framework suitable for widespread use in clinical applications, especially in low-income regions. We plan to release the framework as a part of a free clinical brain imaging analysis tool. The code for this framework is publicly available:https://github.com/Nima-Hs/BraTS.