Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge

Asbjørn Munk; Stefano Cerri; Vardan Nersesjan; Christian Hedeager Krag; Jakob Ambsdorf; Pablo Rocamora García; Julia Machnio; Peirong Liu

cs.CVApr 13, 2026

Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge

Authors: Asbjørn Munk, Stefano Cerri, Vardan Nersesjan, Christian Hedeager Krag, Jakob Ambsdorf, Pablo Rocamora García, Julia Machnio, Peirong Liu, +76 more

Abstract

Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. Self-supervised learning (SSL) can address this by leveraging the vast amounts of unlabeled data produced in clinical workflows to train robust foundation models that adapt out-of-domain with minimal supervision. However, the development of foundation models for brain MRI has been limited by small pretraining datasets and in-domain benchmarking focused on high-quality, research-grade data. To address this gap, we organized the FOMO25 challenge as a satellite event at MICCAI 2025. FOMO25 provided participants with a large pretraining dataset, FOMO60K, and evaluated models on data sourced directly from clinical workflows in few-shot and out-of-domain settings. Tasks covered infarct classification, meningioma segmentation, and brain age regression, and considered both models trained on FOMO60K (method track) and any data (open track). Nineteen foundation models from sixteen teams were evaluated using a standardized containerized pipeline. Results show that (a) self-supervised pretraining improves generalization on clinical data under domain shift, with the strongest models trained out-of-domain surpassing supervised baselines trained in-domain. (b) No single pretraining objective benefits all tasks: MAE favors segmentation, hybrid reconstruction-contrastive objectives favor classification, and (c) strong performance was achieved by small pretrained models, and improvements from scaling model size and training duration did not yield reliable benefits.

Figures & tables

Figure 1: Self-supervised pretraining boosts generalization. Across tasks, the top pretraining-based model from the Method Track outperforms both from-scratch out-of-domain and in-domain supervised baselines, demonstrating that SSL can effectively leverage heterogeneous MRI data. Baselines are nnU-Net (segmentation) and Asparagus (classification/regression). The best performing method for classification was ashash, for segmentation Dolphin_creators, and for regression FOMO2JOMO. For Task 2, “Easy cases” were defined as the top 25% of cases ranked by average Dice score across all challenge participants.

		Overall	Task 1 (Infarct)		Task 2 (Meningioma)			Task 3 (Brain age)
Place	Team	Rank	R	AUROC	R	Dice	NSD	R	Corr.	Mean AE
Method Track
1	FOMO2JOMO	2.3	2	0.782	4	0.095	0.072	1	0.651	10.96
2	Dolphins_creators	4.3	6	0.684	1	0.261	0.232	6	0.448	12.67
3	MaastrichtU-CDS	4.7	5	0.688	2.5	0.181	0.126	6.5	0.599	16.17
4	TUKE	4.8	4	0.750	7	0.046	0.043	3.5	0.602	11.96

Table 1: FOMO25 leaderboard results.

Figure 2: Effect of pretraining choices on downstream performance. (A–C) Pairwise rank differences between tasks, grouped by SSL objective category (global, hybrid, local): classification vs. segmentation (A), segmentation vs. regression (B), and classification vs. regression (C). Teams with Dice or NSD < 0.01 were excluded from (A) and (B). Positive values indicate better relative performance on the task shown at the top of the axis. (D) Overall rank as a function of encoder size. One outlier with 0.5B params was removed from the plot. (E) Overall rank as a function of the number of samples seen during pretraining. (F) Overall rank by whether sequences were treated as i.i.d. during pretraining. (G–I) Per-method rank comparisons across task pairs: classification vs. segmentation (G), classification vs. regression (H), and segmentation vs. regression (I), with lines colored by track.

Figure 3: Comparison of Method and Open Track Submissions. (A) Distribution of pretraining dataset sizes in the Open Track (Method Track teams all used FOMO60K). (B) Dimensionality of the input representation (2D vs. 3D). (C–F) Share of submissions within each track by SSL objective category (global, hybrid, local) (C), encoder size (D), backbone architecture (E), and number of available GPUs (F)

Figure 4: Effect of augmentation strategies on task performance. (A) Task-specific rank as a function of the total number of pretraining augmentations. (B) Task-specific rank as a function of the number of spatial augmentations only. Marker shapes denote track, and colors indicate overall rank within each track. Lower rank indicates better performance. Trend lines are shown for each task.

Figure 5: Impact of hyperparameter tuning on final rank. Categorization of the extent of hyperparameter tuning done for pretraining hyperparameters compared to the final rank on each task.

Figure 6: Rankings with supervised baseline models. (A, B) Bootstrap rank distributions (mean $\pm$ 95% CI) for the Method and Open Tracks, with teams grouped into statistically defined performance tiers (colours). (C-E) Task-level performance for the best out-of-domain method across both tracks alongside in-domain and out-of-domain supervised baselines: (C) ROC curves for the infarct classification task, (D) Dice and NSD score distributions for the meningioma segmentation task, (E) absolute error distributions for the brain-age regression task, with Mean AE and Corr. annotated. Rankings per individual task, along with permutation-test p-value matrices, are provided in Figure 7 for the Method Track and in Figure 8 for the Open Track.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Split	# Cases	Data Origin	Age (mean $\pm$ SD)	Sex (M/F)	Sequences	Slice thickness (mean $\pm$ SD)
Task 1 - Infarct Classification	Finetuning	21	India	61.52 $\pm$ 16.27	16 / 5	FLAIR, DWI (ADC+b1000), T2*(5) / SWI(16)	4.50 $\pm$ 0.89
	Validation	80	Denmark	63.17 $\pm$ 17.27	39 / 41	FLAIR, DWI (ADC+b1000), T2*(15) / SWI(65)	4.57 $\pm$ 1.73
	Test	320	Denmark	64.30 $\pm$ 17.37	161 / 159	FLAIR, DWI (ADC+b1000), T2*(59) / SWI(261)	4.57 $\pm$ 1.77
Task 2 - Meningioma Segmentation	Finetuning	23	India	48.27 $\pm$ 22.04	8 / 15	FLAIR, DWI (b1000), T2*(15) / SWI(8)	4.61 $\pm$ 0.78
	Validation	25	Denmark	65.36 $\pm$ 13.54	7 / 18	FLAIR, DWI (b1000), T2*(2) / SWI(23)	4.11 $\pm$ 2.04
	Test	107	Denmark	67.01 $\pm$ 13.19	30 / 77	FLAIR, DWI (b1000), T2*(16) / SWI(91)	4.28 $\pm$ 2.03

Appendix

Table 2: Summary of datasets used in FOMO25. For each task, we report the split, number of cases, data origin, age, sex, sequences included, and slice thickness. Numbers in parentheses indicate cases per sequence type for T2*/SWI. Slice thickness reported for original data before co-registration, averaged across all sequences.

Figure 7: Statistical analysis for the Method Track. Top row: bootstrap rank distributions for the combined ranking and each individual task, with independently computed performance tiers. Bottom row: pairwise permutation test p-value matrices (lower triangular); asterisks denote statistically significant differences (p < 0.05). Teams are ordered by mean rank within each task.

Figure 8: Statistical analysis for the Open Track. Top row: bootstrap rank distributions for the combined ranking and each individual task, with independently computed performance tiers. Bottom row: pairwise permutation test p-value matrices (lower triangular); asterisks denote statistically significant differences (p < 0.05). Teams are ordered by mean rank within each task.

Figure 9: Overview of submitted methods . (A) Pretraining augmentations used across all submissions, grouped by spatial, intensity, and artifact categories and colored by SSL objective type (Global, Hybrid, Local). (B) Distribution of GPU types used for pretraining. (C) Backbone architectures, showing the proportion of CNN, Swin Transformer, Uniformer, Vision Transformer, and xLSTM models. (D) Count of SSL objective types by category. (E) Number of GPUs used during pretraining. (F) Pretraining patch sizes, stratified by track. (G) Encoder parameter counts. Panels E–G separate submissions per track.

Explore similar work

BrainNext: A General-Purpose Self-Supervised Foundation Model for Brain MRI Analysis
Jul 20, 2026Moona Mazher, Abdul Qayyum, Steven A. Niederer +1Medical Imaging Foundation Models Self-Supervised Pre-Training
UCL Hawkes Institute, Department of Computer Science, University College London, London, United Kingdom · National Heart & Lung Institute, Imperial College London, London, United Kingdom · Division of Cardiovascular Medicine, Stanford Cardiovascular Institute, Stanford University, California, USA
BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning
Apr 30, 2026Yizhou Wu, Shansong Wang, Yuheng Li +5Medical Imaging Foundation Models Self-Supervised Pre-Training
Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA, USA. · Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, USA. · Department of Biomedical Engineering, Georgia Institute of Technology, Atlanta, GA, USA. +2
Combining General and Domain-Specific Pretext Tasks for Brain MR Image Segmentation
Sep 25, 2026Tasneem Nasser, Susanne Schmid, Roberto Souza +1Medical Image Representation Learning Self-Supervised Pre-Training
Biomedical Engineering University of Calgary Calgary AB Canada · Hotchkiss Brain Institute University of Calgary Calgary AB Canada · Electrical and Software Engineering University of Calgary Calgary AB Canada +1

Jul 20, 2026J/K move · Enter open · S save

BrainNext: A General-Purpose Self-Supervised Foundation Model for Brain MRI Analysis

Moona Mazher, Abdul Qayyum, Steven A. Niederer, Daniel C. Alexander

UCL Hawkes Institute, Department of Computer Science, University College London, London, United Kingdom · National Heart & Lung Institute, Imperial College London, London, United Kingdom · Division of Cardiovascular Medicine, Stanford Cardiovascular Institute, Stanford University, California, USA

Medical Imaging Foundation Models Self-Supervised Pre-Training Medical Image Representation Learning

Open paper AskPDF arXiv

Foundation models pretrained using self-supervised learning have transformed computer vision by learning transferable representations from large-scale unlabeled data. However, existing foundation models for neuroimaging remain limited by task-specific training, slice-based learning strategies, or relatively small pretraining datasets, restricting their generalizability across diverse brain MRI applications. In this work, we present BrainNext, a general-purpose self-supervised foundation model for volumetric brain MRI analysis. BrainNext combines masked autoencoder (MAE) pretraining with a native three-dimensional Bi-Directional xLSTM-UNet architecture to learn rich anatomical representations from 60,551 unlabeled brain MRI examinations spanning multiple MRI modalities. The pretrained model is subsequently adapted to downstream tasks through lightweight task-specific fine-tuning. We evaluate BrainNext on the Foundation Models for Medical Imaging (FOMO) 2025 Method Track, encompassing classification, segmentation, and brain-age estimation, where it achieved second place overall and ranked first in the meningioma segmentation task on the official FOMO 2025 challenge leaderboard, demonstrating strong transferability across heterogeneous neuroimaging tasks. These results highlight the potential of large-scale self-supervised pretraining to learn robust and transferable volumetric representations, establishing BrainNext as a scalable foundation model for diverse brain MRI applications.

Abstract

Explore similar work

BrainNext: A General-Purpose Self-Supervised Foundation Model for Brain MRI Analysis

BrainDINO: A Brain MRI Foundation Model for Generalizable Clinical Representation Learning

Combining General and Domain-Specific Pretext Tasks for Brain MR Image Segmentation