Multi-modal U-Nets with Boundary Loss and Pre-training for Brain Tumor Segmentation.

Gliomas are the most common primary brain tumors, and their manual segmentation is a time-consuming and user-dependent process. We present a two-step multi-modal U-Net-based architecture with unsupervised pre-training and surface loss component for brain tumor segmentation which allows us to seamlessly benefit from all magnetic resonance modalities during the delineation. The results of the experimental study, performed over the newest release of the BraTS test set, revealed that our method delivers accurate brain tumor segmentation, with the average DICE score of 0.72, 0.86, and 0.77 for the enhancing tumor, whole tumor, and tumor core, respectively. The total time required to process one study using our approach amounts to around 20 s.