Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge
Researchers organized the FOMO25 challenge to address the difficulties in deploying AI for brain MRI analysis in clinical settings. The challenge provided a large pretraining dataset, FOMO60K, and evaluated nineteen foundation models on tasks like infarct classification and brain age regression using data from actual clinical workflows. Findings indicated that self-supervised pretraining enhances generalization on clinical data, with some out-of-domain trained models outperforming in-domain supervised models. The study also noted that different pretraining objectives are better suited for specific tasks, and that smaller pretrained models can achieve strong performance without significant benefits from scaling. AI
IMPACT Highlights the potential of self-supervised learning and foundation models to improve AI's generalization in clinical brain MRI analysis.