Addressing Domain Shift in Low-Resource Neuroimaging: Glioma Segmentation for Sub-Saharan MRI

Accurate segmentation of gliomas from magnetic resonance imaging (MRI) is crucial for diagnosis and treatment planning. However, models trained on high-resource datasets degrade sharply when applied to Sub-Saharan African (SSA) scans, which often exhibit heterogeneous acquisition protocols, reduced resolution, and limited expert annotations. This domain gap has reinforced inequities in neuro-oncology outcomes, as automated tools remain largely unavailable in regions where they are most urgently needed. We propose a domain-adaptive transformer framework that integrates intensity harmonization with architectural refinements for robust performance in SSA neuroimaging. Our method builds on a SegFormer-based volumetric backbone, augmented with a wavelet-convolutional input stem for frequency-aware encoding and dual attention modules for spatial–channel refinement. Decoder outputs are guided through radiomics-driven stratification to improve boundary delineation under low-contrast conditions. To mitigate domain shift, we apply histogram matching between SSA and BraTS 2023 adult glioma scans, followed by pretraining on BraTS 2023 and fine-tuning on BraTS-Africa. The resulting model has only 12M parameters and achieves 2.3× faster inference than CNN baselines, supporting use in resource-constrained environments. Datasets. BraTS 2023 comprises 1,251 annotated adult glioma cases across four MRI modalities. BraTS-Africa extends this benchmark to 60 labeled and 35 validation cases from SSA institutions, capturing real-world challenges such as lower field strength and variable imaging protocols. Results. On BraTS-Africa, our model achieves mean Dice of 0.842 (vs. 0.589 for SSA-only training and 0.681 for direct transfer), reducing mean HD95 from 26.27 mm to 7.12 mm. Source-domain performance is preserved at 0.891 Dice, showing effective transfer without catastrophic forgetting. Ablations show histogram matching contributes +0.089 Dice and pretraining adds +0.164, with the largest gains in enhancing tumors. The model surpasses recent domain adaptation baselines, including CORAL (+0.133 Dice) and domain-adversarial training (+0.147 Dice). Clinical Significance. Performance (0.842 Dice) exceeds inter-rater agreement on BraTS-Africa (0.78–0.82 Dice), suggesting clinical viability. This work highlights how domain adaptation can deliver equitable neuroimaging solutions for underrepresented populations. Future efforts will expand to multi-institutional SSA datasets and generalize to other neurological conditions common in low-resource settings.