Brain Tumor Segmentation Based on Self-supervised Pre-training and Adaptive Region-Specific Loss

Automatic segmentation of pediatric brain tumors has garnered increasing attention due to its crucial role in diagnosis and treatment planning, which is challenging due to tumors’ rarity and heterogeneity. This work presents a high-performance segmentation method based on nnU-Net for the pediatric brain tumor segmentation (BraTS-PEDs) challenge 2023. Firstly, we leverage the BraTS-PEDs 2023 dataset to pre-train our network with self-supervised learning strategies that encourage the network to grasp both global and local features inherent to pediatric brain tumors. Secondly, we leverage an adaptive region-specific training loss to tackle the challenges posed by the heterogeneity of magnetic resonance imaging (MRI) and the imbalance between different tissue categories during training, which effectively enhances the network’s ability to learn from difficult regions. Our approach achieves lesion-wise Dice scores of 82.68%, 74.06%, and 59.63% with corresponding lesion-wise HD95 (mm) values of 22.85, 29.22, and 121.02 for the whole tumor (WT), non-enhancing component (NC), and enhancing tumor (ET) respectively, as determined by the online evaluation platform. We also validate our method on the BraTS meningioma challenge 2023 at the online evaluation platform and get lesion-wise Dice scores of 79.69%, 80.61%, and 79.97% with corresponding lesion-wise HD95 (mm) values of 49.37, 47.18, and 47.38 for the non-enhancing tumor core (NT), edema (ED), and ET.