Medical AI models fail on African brain data study finds
New research reveals medical AI models trained on Western data underperform on African brain scans
A new study published on arXiv examines whether medical foundation models (FMs) effectively generalize to African brain MRI data, finding that current challenges stem from dataset limitations rather than inherent model bias.
Researchers from Erasmus MC and other institutions tested four FMs—BrainIAC, 3DINO, MedSAM2, and Medical-SAM2—on two African datasets: a Nigerian dementia classification task and BraTS-Africa for brain tumor segmentation. While models like MedSAM2 achieved a Dice score of 0.86 in segmentation, gains were inconsistent across tasks. The team found no clear geographic bias, attributing performance disparities instead to the limited size and diversity of African neuroimaging datasets. The findings were presented at the AFRICAI workshop, co-located with MICCAI 2026.
- Four medical FMs (BrainIAC, 3DINO, MedSAM2, Medical-SAM2) evaluated on Nigerian and BraTS-Africa datasets
- Highest ROC-AUC for dementia classification: 0.86 (BrainIAC); best Dice score: 0.86 (MedSAM2 for segmentation)
- Performance gaps linked to dataset scarcity, not geographic bias
Why It Matters
Exposes critical data gaps threatening equitable medical AI deployment in underrepresented regions