FOMO26: Foundation Model Challenge for Brain MRI

Large-scale self-supervised pretraining offers significant yet underutilised opportunities for advancing brain magnetic resonance imaging (MRI) analysis. Despite recent progress, most state-of-the-art methods continue to rely on fully supervised learning. While strong data augmentation can yield competitive performance with limited annotations, such approaches often struggle under domain shift and degraded clinical image quality, limiting their reliability in real-world settings. Foundation models trained via self-supervised pretraining on large, heterogeneous datasets offer a promising alternative, but their true benefits over supervised models remain insufficiently and inconsistently evaluated.

To address this gap, we present a new edition of FOMO, a MICCAI challenge designed to rigorously assess foundation models for brain MRI using large-scale, diverse datasets and comprehensive evaluation, with a strong emphasis on real-world clinical routine data and out-of-domain settings. A central contribution of this challenge is a substantially expanded pretraining dataset comprising 318,877 brain MRI scans from 82,678 MRI sessions and 59,969 subjects, aggregated from 920 publicly available sources, providing a realistic substrate for large-scale self-supervised representation learning.

This year’s challenge comprises seven downstream tasks, spanning image-level classification, voxel-wise segmentation, regression, and representation quality assessment. The first three tasks continue benchmarks from the previous edition, enabling longitudinal comparison, while four new tasks substantially broaden the clinical and methodological scope. For all tasks, fine-tuning is explicitly limited to few-shot regimes. Evaluation is performed on hidden, large-scale, multi-institutional out-of-domain clinical datasets, comprising several hundred to several thousand cases per task, ensuring that performance reflects genuine generalization rather than task-specific overfitting.

To encourage methodological transparency and broad participation, each task is offered under two complementary tracks: a Method track, restricting participants to challenge-provided data only, and an Open track, permitting the use of external or private data. This dual-track design enables fair comparison of representation learning strategies while also allowing exploration of scaling effects, architectural choices, and data curation strategies beyond the challenge data alone.

By evaluating pretrained models across a diverse set of clinically grounded downstream tasks, the challenge is designed to move beyond task-specific or noise-driven improvements and to enable a more reliable assessment of whether self-supervised pretraining yields consistent and generalizable advantages over fully supervised approaches. Ultimately, FOMO26 aims to advance the development of clinically reliable foundation models for brain MRI, with attention to robustness, fairness, and real-world applicability.