Accurate glioma segmentation is critical for clinical diagnosis and treatment planning, yet remains challenging due to infiltrative tumor growth, heterogeneous imaging protocols, and scarcity of expert annotations. We present MAGPIE, a self-supervised learning framework that combines masked autoencoding, contrastive learning, and sparse mixture of experts to enable accurate glioma segmentation with minimal labeled data. By pretraining on 43,505 unlabeled multi-modal brain MRI scans, MAGPIE learns generalizable representations through a channel-agnostic architecture that handles varying modality configurations without protocol-specific preprocessing. The sparse MoE mechanism with top-2 routing allows specialized expert networks to emerge for different glioma subregions, while deformable attention mechanisms capture infiltrative margins and multi-scale features. Fine-tuning on only 20 labeled cases achieves 60.87% Dice score on BraTS21, a 2.59% absolute improvement over training from scratch, with 70.32% on out-of-distribution data demonstrating robust cross-domain generalization. These results reduce annotation requirements by 95% compared to typical supervised methods, directly addressing the data scarcity bottleneck in rare tumor subtypes and enabling deployment across heterogeneous clinical imaging systems.
Xie et al. (Sat,) studied this question.