Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge
Overview
Deploying automated brain MRI analysis in the clinic is hard: clinical data is noisy and heterogeneous, and high-quality labels are expensive to obtain. Self-supervised learning (SSL) offers a way to leverage the huge volume of unlabeled clinical scans to train foundation models — but progress has been limited by small pretraining datasets and benchmarks that only use clean, research-grade data.
Method
To close this gap, the authors organised FOMO25, a satellite challenge at MICCAI 2025. It released FOMO60K, a large unlabeled pretraining set of over 60,000 structural brain MRI scans from both clinical and research settings, and evaluated submitted models directly on data drawn from clinical workflows, under few-shot and out-of-domain conditions across tasks including infarct classification, meningioma segmentation, and brain age regression.
Results
The paper reports the challenge design, the FOMO60K dataset, and the findings across participating methods, offering a benchmark and set of lessons for building brain MRI foundation models that generalise to messy, real-world clinical settings rather than just clean research data.