Overcoming Data Scarcity: Brain Tumor Segmentation in Pediatric and African Populations

Abstract

Deep learning has enabled significant progress in brain tumor segmentation, particularly for adult gliomas using large curated and annotated datasets. However, models often underperform on underrepresented populations such as pediatric and African cohorts due to limited annotated data. Leveraging large adult datasets in mixed training settings can introduce domain bias, due to differences in tumor location, morphology, and imaging quality. To address this, we propose a meta-learning-based adaptive reweighting framework that estimates gradient alignment between adult samples and a small subset of pediatric or African cases. Samples with higher alignment receive greater weight during training, guiding the model toward domain-relevant features and suppressing misleading adult-specific patterns. Experiments on BraTS-PEDs and BraTS-Africa demonstrate consistent performance gains over baseline and prior reweighting methods. The code will be available upon acceptance.