arxiv: 2604.11679 · v1 · submitted 2026-04-13 · 💻 cs.CV

Recognition: unknown

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

Asbj{\o}rn Munk , Stefano Cerri , Vardan Nersesjan , Christian Hedeager Krag , Jakob Ambsdorf , Pablo Rocamora Garc\'ia , Julia Machnio , Peirong Liu

show 76 more authors

Suhyun Ahn Nasrin Akbari Yasmina Al Khalil Kimberly Amador Sina Amirrajab Tal Arbel Meritxell Bach Cuadra Ujjwal Baid Bhakti Baheti Jaume Banus Kamil Barbierik Christoph Brune Yansong Bu Baptiste Callard Yuhan Chen Cornelius Crijnen Corentin Dancette Peter Drotar Prasad Dutande Nils D. Forkert Saurabh Garg Jakub Gazda Matej Gazda Beno\^it G\'erin Partha Ghosh Weikang Gong Pedro M. Gordaliza Sam Hashemi Tobias Heimann Fucang Jia Jiexin Jiang Emily Kaczmarek Chris Kang Seung Kwan Kang Mohammad Khazaei Julien Khlaut Petros Koutsouvelis Jae Sung Lee Yuchong Li Mengye Lyu Mingchen Ma Anant Madabhushi Klaus H. Maier-Hein Pierre Manceron Andr\'es Mart\'inez Mora Moona Mazher Felix Meister Nataliia Molchanova Steven A. Niederer Leonard N\"urnberg Jinah Park Abdul Qayyum Jonas Richiardi Antoine Saporta Branislav Setlak Ning Shen Justin Szeto Constantin Ulrich Puru Vaish Vibujithan Vigneshwaran Leroy Volmer Zihao Wang Siqi Wei Anthony Winder Jelmer M. Wolterink Maxence Wynen Chang Yang Si Young Yie Mostafa Mehdipour Ghazi Akshay Pai Espen Jimenez Solem Sebastian N{\o}rgaard Llambias Mikael Boesen Michael Eriksen Benros Juan Eugenio Iglesias Mads Nielsen

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:41 UTC · model grok-4.3

classification 💻 cs.CV

keywords brain MRIself-supervised learningfoundation modelsclinical datadomain shiftFOMO25 challengefew-shot learninggeneralization

0 comments

The pith

Self-supervised pretraining on large unlabeled brain MRI data improves generalization to noisy clinical scans over supervised in-domain training.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper runs the FOMO25 challenge to test foundation models for brain MRI analysis on data taken straight from clinical workflows. It supplies a large unlabeled pretraining set called FOMO60K and asks teams to build models that then handle three tasks in few-shot out-of-domain conditions: infarct classification, meningioma segmentation, and brain age regression. Results from nineteen models show that the strongest self-supervised models trained outside the target domain outperform supervised models trained directly on the clinical labels. The work also finds that different pretraining objectives suit different tasks and that small models often perform as well as larger ones.

Core claim

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. No single pretraining objective benefits all tasks: MAE favors segmentation while hybrid reconstruction-contrastive objectives favor classification. Strong performance was achieved by small pretrained models, and improvements from scaling model size and training duration did not yield reliable benefits.

What carries the argument

The FOMO25 challenge evaluation pipeline, which uses the FOMO60K unlabeled pretraining dataset and standardized containerized testing on clinical workflow data splits for the three target tasks.

If this is right

Self-supervised pretraining enables better performance when models encounter clinical data that differs from their training distribution.
MAE-style pretraining supports segmentation tasks while hybrid objectives support classification tasks.
Small pretrained models can match or exceed larger ones on these clinical tasks without further scaling.
Foundation models can reduce reliance on costly new labels collected for each hospital's specific data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Hospitals could deploy a single pretrained model across varied scanners with only minimal new labels per site.
Task-specific pretraining choices might be tuned in advance to match common clinical needs such as tumor outlining or stroke detection.
The same approach could be extended to other modalities like CT if similar large unlabeled clinical archives become available.

Load-bearing premise

The FOMO60K pretraining dataset and the selected clinical tasks and data splits capture enough of the real heterogeneity and noise found in everyday hospital brain MRI scans.

What would settle it

A follow-up test on brain MRI scans from a new group of hospitals or scanner vendors where the top FOMO25 self-supervised models no longer outperform supervised baselines trained on those new scans.

Figures

Figures reproduced from arXiv: 2604.11679 by Abdul Qayyum, Akshay Pai, Anant Madabhushi, Andr\'es Mart\'inez Mora, Anthony Winder, Antoine Saporta, Asbj{\o}rn Munk, Baptiste Callard, Beno\^it G\'erin, Bhakti Baheti, Branislav Setlak, Chang Yang, Chris Kang, Christian Hedeager Krag, Christoph Brune, Constantin Ulrich, Corentin Dancette, Cornelius Crijnen, Emily Kaczmarek, Espen Jimenez Solem, Felix Meister, Fucang Jia, Jae Sung Lee, Jakob Ambsdorf, Jakub Gazda, Jaume Banus, Jelmer M. Wolterink, Jiexin Jiang, Jinah Park, Jonas Richiardi, Juan Eugenio Iglesias, Julia Machnio, Julien Khlaut, Justin Szeto, Kamil Barbierik, Kimberly Amador, Klaus H. Maier-Hein, Leonard N\"urnberg, Leroy Volmer, Mads Nielsen, Matej Gazda, Maxence Wynen, Mengye Lyu, Meritxell Bach Cuadra, Michael Eriksen Benros, Mikael Boesen, Mingchen Ma, Mohammad Khazaei, Moona Mazher, Mostafa Mehdipour Ghazi, Nasrin Akbari, Nataliia Molchanova, Nils D. Forkert, Ning Shen, Pablo Rocamora Garc\'ia, Partha Ghosh, Pedro M. Gordaliza, Peirong Liu, Peter Drotar, Petros Koutsouvelis, Pierre Manceron, Prasad Dutande, Puru Vaish, Sam Hashemi, Saurabh Garg, Sebastian N{\o}rgaard Llambias, Seung Kwan Kang, Sina Amirrajab, Siqi Wei, Si Young Yie, Stefano Cerri, Steven A. Niederer, Suhyun Ahn, Tal Arbel, Tobias Heimann, Ujjwal Baid, Vardan Nersesjan, Vibujithan Vigneshwaran, Weikang Gong, Yansong Bu, Yasmina Al Khalil, Yuchong Li, Yuhan Chen, Zihao Wang.

**Figure 1.** 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, … view at source ↗

**Figure 2.** 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 t… view at source ↗

**Figure 3.** 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) three task-specific… view at source ↗

**Figure 4.** 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. cash prize of $2000, with $… view at source ↗

**Figure 5.** 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. between ranks over task types grouped by SSL objective type. We find that local objectives favor segmentation over classification, with no notable difference between either segmentation and regression or classification and regres… view at source ↗

**Figure 6.** Figure 6: Rankings with supervised baseline models. (A, B) Bootstrap rank distributions (mean ± 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 N… view at source ↗

read the original 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 \textit{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 \textit{out-of-domain} surpassing supervised baselines trained \textit{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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SSL pretraining on FOMO60K beats in-domain supervised baselines on the three clinical tasks, but the paper gives limited evidence that the chosen data and splits reflect typical clinical noise and heterogeneity.

read the letter

The main result is that self-supervised models trained out-of-domain on the large FOMO60K set outperform supervised models trained in-domain on infarct classification, meningioma segmentation, and brain age regression, with different pretraining objectives suiting different tasks and little consistent gain from scaling model size or training time. The standardized containerized evaluation across 19 models from 16 teams makes the head-to-head comparison credible and directly addresses the data heterogeneity problem in clinical brain MRI. That setup is the paper's clearest contribution: it moves beyond research-grade benchmarks to few-shot and out-of-domain clinical data in a controlled way. The finding that small pretrained models already perform strongly is also useful practical information. The soft spot is the stress-test concern about representativeness. The abstract states the evaluation data come from clinical workflows, yet the provided details do not include scanner overlap statistics, artifact rates, intensity shifts, or cohort differences that would show how severe the domain shift actually is. If the selected tasks and splits are cleaner or the shift milder than average clinical cases, the generalization claim does not fully support the broader framing of foundation models ready for the clinic. The paper is aimed at researchers working on self-supervised learning and foundation models for medical imaging who need benchmark numbers on pretraining objectives and scaling under domain shift. It deserves a serious referee because the experimental design is reproducible and the questions it asks are the right ones for deployment, even though additional quantitative characterization of the domain shifts would make the results more convincing.

Referee Report

2 major / 2 minor

Summary. The manuscript reports results from the FOMO25 challenge, which provides the FOMO60K pretraining dataset for self-supervised learning of brain MRI foundation models. Nineteen models from sixteen teams are evaluated via a standardized containerized pipeline on three tasks using data from clinical workflows in few-shot and out-of-domain settings: infarct classification, meningioma segmentation, and brain age regression. Key claims are that self-supervised pretraining improves generalization under domain shift (with strongest out-of-domain SSL models surpassing in-domain supervised baselines), that no single pretraining objective benefits all tasks (MAE favors segmentation while hybrid objectives favor classification), and that small models achieve strong performance without reliable gains from scaling model size or training duration.

Significance. If the empirical results hold under the chosen data, this provides a controlled multi-team benchmark supporting the utility of large-scale SSL pretraining on heterogeneous unlabeled data for clinical brain MRI tasks. The standardized evaluation and explicit comparison of method-track (FOMO60K only) versus open-track models offer concrete evidence on generalization benefits and objective-task interactions that can guide future foundation model development.

major comments (2)

Abstract: The central claim that out-of-domain SSL models surpass in-domain supervised baselines and that SSL improves generalization on clinical data under domain shift rests on the assumption that the FOMO60K pretraining set and the three evaluation tasks exhibit representative clinical heterogeneity and noise. The abstract states evaluation uses 'data sourced directly from clinical workflows' yet provides no quantitative metrics of domain shift (scanner metadata overlap, intensity distribution statistics, artifact prevalence, or patient cohort differences) between pretraining and evaluation data; without this, the broader 'foundation models for the clinic' framing is not fully supported by the presented evidence.
Results section (findings a and b): The reported superiority of specific pretraining objectives for particular tasks (MAE for segmentation, hybrid for classification) and the overall generalization benefit lack accompanying statistical details such as confidence intervals, p-values from paired tests across teams, or effect sizes. Given the multi-team setup and potential variability in containerized runs, these omissions make it difficult to assess whether the observed differences are robust or task-specific artifacts.

minor comments (2)

The description of the standardized containerized evaluation pipeline would be strengthened by an explicit reference to the public code repository or a supplementary table listing exact preprocessing steps, few-shot sample counts, and cross-validation folds used for each task.
A summary table of the top-performing models' pretraining objectives, model sizes, and training durations (method track vs. open track) would improve readability and allow readers to directly map the scaling observations to specific entries.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our FOMO25 challenge manuscript. The comments highlight opportunities to strengthen the evidence for domain shift and the statistical robustness of our findings. We address each major comment below, indicating where revisions will be made to the manuscript.

read point-by-point responses

Referee: Abstract: The central claim that out-of-domain SSL models surpass in-domain supervised baselines and that SSL improves generalization on clinical data under domain shift rests on the assumption that the FOMO60K pretraining set and the three evaluation tasks exhibit representative clinical heterogeneity and noise. The abstract states evaluation uses 'data sourced directly from clinical workflows' yet provides no quantitative metrics of domain shift (scanner metadata overlap, intensity distribution statistics, artifact prevalence, or patient cohort differences) between pretraining and evaluation data; without this, the broader 'foundation models for the clinic' framing is not fully supported by the presented evidence.

Authors: We agree that explicit quantitative metrics of domain shift would strengthen the manuscript and better support the clinical foundation model framing. The challenge design operationalizes out-of-domain evaluation through data sourced from distinct clinical workflows and sites, with performance gains providing supporting evidence. In the revised version, we will add a dedicated paragraph and table in the Methods or Results section summarizing available metadata (scanner vendors, field strengths, and basic intensity distribution statistics) between FOMO60K and the evaluation sets. Where full metadata is unavailable due to anonymization constraints, we will explicitly discuss this limitation and adjust the abstract wording to more precisely reflect the operational definition of domain shift used in the challenge. revision: yes
Referee: Results section (findings a and b): The reported superiority of specific pretraining objectives for particular tasks (MAE for segmentation, hybrid for classification) and the overall generalization benefit lack accompanying statistical details such as confidence intervals, p-values from paired tests across teams, or effect sizes. Given the multi-team setup and potential variability in containerized runs, these omissions make it difficult to assess whether the observed differences are robust or task-specific artifacts.

Authors: We concur that additional statistical details are necessary to demonstrate robustness, particularly given the multi-team and containerized evaluation setup. In the revised manuscript, we will report 95% confidence intervals for all primary metrics (computed via bootstrapping across the few-shot splits). We will also add paired statistical comparisons (Wilcoxon signed-rank tests) between the top SSL models and the in-domain supervised baselines, including p-values and effect sizes. These will be incorporated into the Results text, tables, and figure captions for findings (a) and (b). revision: yes

Circularity Check

0 steps flagged

No significant circularity: purely empirical benchmark study

full rationale

The paper reports results from the FOMO25 challenge comparing self-supervised pretraining on FOMO60K against supervised baselines on three clinical tasks (infarct classification, meningioma segmentation, brain age regression). All claims derive directly from standardized containerized evaluations on the provided data splits; there are no equations, fitted parameters renamed as predictions, self-definitional constructs, or load-bearing self-citations that reduce the central generalization result to inputs fitted within the paper. The study is self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Empirical challenge report with no mathematical model, free parameters, or postulated entities; relies on standard assumptions of SSL and supervised learning.

pith-pipeline@v0.9.0 · 6005 in / 952 out tokens · 52668 ms · 2026-05-10T16:41:30.144342+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Beyond Instance-Level Self-Supervision in 3D Multi-Modal Medical Imaging
cs.CV 2026-05 unverdicted novelty 5.0

A self-supervised approach uses consistent spatial relationships of anatomical structures across patients to improve 3D multi-modal medical image representations, yielding modest gains on segmentation and classificati...

Reference graph

Works this paper leans on

54 extracted references · 16 canonical work pages · cited by 1 Pith paper · 9 internal anchors

[1]

Isensee, P

F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, K. H. Maier-Hein, nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation, Nature methods 18 (2021) 203–211

2021
[2]

Isensee, T

F. Isensee, T. Wald, C. Ulrich, M. Baumgartner, S. Roy, K. Maier-Hein, P. F. Jaeger, nnu-net revisited: A call for rigorous validation in 3d medical image segmentation, in: International Conference on Medical Image Comput- ing and Computer-Assisted Intervention, Springer, 2024, pp. 488–498

2024
[3]

Mårtensson, D

G. Mårtensson, D. Ferreira, T. Granberg, L. Cavallin, K. Oppedal, A. Padovani, I. Rektorova, L. Bonanni, M. Pardini, M. G. Kramberger, J.-P. Taylor, J. Hort, J. Snædal, J. Kulisevsky, F. Blanc, A. Antonini, P. Mecocci, B. Vellas, M. Tsolaki, I. Kłoszewska, H. Soininen, S. Lovestone, A. Simmons, D. Aarsland, E. Westman, The reliability of a deep learning m...

work page doi:10.1016/j.media.2020.101714 2020
[4]

E. A. AlBadawy, A. Saha, M. A. Mazurowski, Deep learn- ing for segmentation of brain tumors: Impact of cross- institutional training and testing, Medical physics 45 (2018) 1150–1158

2018
[5]

Nørgaard Llambias, M

S. Nørgaard Llambias, M. Nielsen, M. Mehdipour Ghazi, Data augmentation-based unsupervised domain adapta- tion in medical imaging, in: Scandinavian Conference on Image Analysis, Springer, 2025, pp. 177–186

2025
[6]

Smith-Bindman, D

R. Smith-Bindman, D. L. Miglioretti, E. Johnson, C. Lee, H. S. Feigelson, M. Flynn, R. T. Greenlee, R. L. Kruger, M. C. Hornbrook, D. Roblin, et al., Use of diagnos- tic imaging studies and associated radiation exposure for patients enrolled in large integrated health care systems, 1996-2010, Jama 307 (2012) 2400–2409

1996
[7]

Smith-Bindman, M

R. Smith-Bindman, M. L. Kwan, E. C. Marlow, M. K. Theis, W. Bolch, S. Y . Cheng, E. J. Bowles, J. R. Duncan, R. T. Greenlee, L. H. Kushi, et al., Trends in use of med- ical imaging in US health care systems and in Ontario, Canada, 2000-2016, Jama 322 (2019) 843–856. 13

2000
[8]

Devlin, M.-W

J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre- training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 conference of the North American chapter of the association for com- putational linguistics: human language technologies, vol- ume 1 (long and short papers), 2019, pp. 4171–4186

2019
[9]

T. Chen, S. Kornblith, M. Norouzi, G. Hinton, A simple framework for contrastive learning of visual representa- tions, in: International conference on machine learning, PmLR, 2020, pp. 1597–1607

2020
[10]

DINOv2: Learning Robust Visual Features without Supervision

M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haz- iza, F. Massa, A. El-Nouby, et al., DINOv2: Learning Robust Visual Features without Supervision, arXiv preprint arXiv:2304.07193 (2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[11]

Y . Shi, I. Daunhawer, J. E. V ogt, P. Torr, A. Sanyal, How robust is unsupervised representation learning to distri- bution shift?, in: The Eleventh International Confer- ence on Learning Representations, 2023. URL:https: //openreview.net/forum?id=LiXDW7CF94J

2023
[12]

K. He, X. Chen, S. Xie, Y . Li, P. Dollár, R. Girshick, Masked autoencoders are scalable vision learners, in: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16000–16009

2022
[13]

Brown, B

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Ka- plan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learn- ers, Advances in neural information processing systems 33 (2020) 1877–1901

2020
[14]

Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V . Stoyanov, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019)

work page internal anchor Pith review Pith/arXiv arXiv 1907
[15]

Y . Tang, D. Yang, W. Li, H. R. Roth, B. Landman, D. Xu, V . Nath, A. Hatamizadeh, Self-supervised pre-training of swin transformers for 3d medical image analysis, in: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 20730–20740

2022
[16]

Z. Chen, D. Agarwal, K. Aggarwal, W. Safta, M. M. Balan, K. Brown, Masked image modeling advances 3d medical image analysis, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 1970–1980

2023
[17]

L. Wu, J. Zhuang, H. Chen, V oco: A simple-yet-effective volume contrastive learning framework for 3d medical im- age analysis, in: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, 2024, pp. 22873–22882

2024
[18]

T. Wald, C. Ulrich, S. Lukyanenko, A. Goncharov, A. Paderno, M. Miller, L. Maerkisch, P. Jaeger, K. Maier- Hein, Revisiting mae pre-training for 3d medical im- age segmentation, in: Proceedings of the Computer Vi- sion and Pattern Recognition Conference, 2025, pp. 5186– 5196

2025
[19]

A large-scale heterogeneous 3D magnetic resonance brain imaging dataset for self-supervised learning

S. Cerri, A. Munk, S. N. Llambias, J. Ambsdorf, J. Machnio, V . Nersesjan, C. Hedeager Krag, P. Liu, P. Rocamora García, M. Mehdipour Ghazi, M. Boesen, M. E. Benros, J. E. Iglesias, M. Nielsen, A large- scale heterogeneous 3D magnetic resonance brain imag- ing dataset for self-supervised learning, arXiv preprint arXiv:2506.14432 (2026). URL:https://arxiv....

work page internal anchor Pith review Pith/arXiv arXiv 2026
[20]

Ulrich, T

C. Ulrich, T. Wald, Y . Kirchhoff, M. Knopp, R. Peret- zke, M. Fischer, P. Ghosh, F. Isensee, A. Hilbert, P. Naser, L. Wessel, M. Foltyn-Dumitru, G. Brug- nara, J. B. Fiebach, J. O. Neumann, L. König, P. V ollmuth, K. Maier-Hein, SSL3D, 2025.https:// ssl3d-challenge.dkfz.de/home

2025
[21]

T. Wald, C. Ulrich, J. Suprijadi, S. Ziegler, M. Nohel, R. Peretzke, G. Kohler, K. Maier-Hein, An OpenMind for 3D medical vision self-supervised learning, in: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 23839–23879

2025
[22]

J. Ma, Y . Zhou, B. Wang, S. Kim, Z. Marinov, C. Xing, F. Li, Y . He, W. Li, F. Isensee, M. Rokuss, L. Krämer, K. Maier-Hein, Y . Du, B. Zhao, H. Wang, J. He, Y . Qiao, M. Zhang, H. Zhang, G.-Z. Yang, Y . Gu, L. Lumetti, F. Bolelli, C. Grana, Y . Chen, A. Erturk, T. Kuestner, S. Gatidis, M. Ingrisch, R. Graf, H. Möller, J. Kirschke, Z. Lin, T. Tan, H. Qu,...

2025
[23]

J. E. Iglesias, B. Billot, Y . Balbastre, C. Magdamo, S. E. Arnold, S. Das, B. L. Edlow, D. C. Alexander, P. Gol- land, B. Fischl, SynthSR: A public AI tool to turn het- erogeneous clinical brain scans into high-resolution T1- weighted images for 3D morphometry, Science advances 9 (2023) eadd3607

2023
[24]

Cerri, V

S. Cerri, V . Nersesjan, K. V . Klein, E. C. Cóppulo, S. N. Llambias, M. M. Ghazi, M. Nielsen, M. E. Benros, Cross- disorder comparison of brain structures among 4,836 indi- viduals with mental disorders and controls utilizing danish population-based clinical mri scans, Molecular Psychiatry (2026). 14

2026
[25]

A. Munk, J. Ambsdorf, S. Llambias, M. Nielsen, Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation, MICCAI Workshop on Advancing Data Solutions in Medical Imag- ing AI (ADSMI 2024), MICCAI 2024 (2024)

2024
[26]

S. N. Llambias, J. Machnio, A. Munk, J. Ambsdorf, M. Nielsen, M. M. Ghazi, Yucca: A deep learning framework for medical image analysis, arXiv preprint arXiv:2407.19888 (2024)

work page arXiv 2024
[27]

Maier-Hein, A

L. Maier-Hein, A. Reinke, P. Godau, M. D. Tizabi, F. Buettner, E. Christodoulou, B. Glocker, F. Isensee, J. Kleesiek, M. Kozubek, et al., Metrics reloaded: recom- mendations for image analysis validation, Nature methods 21 (2024) 195–212

2024
[28]

Phipson, G

B. Phipson, G. K. Smyth, Permutation P-values should never be zero: calculating exact P-values when permutations are randomly drawn, Statistical appli- cations in genetics and molecular biology 9 (2010). URL:https://doi.org/10.2202/1544-6115.1585. doi:10.2202/1544-6115.1585

work page doi:10.2202/1544-6115.1585 2010
[29]

Maier-Hein, M

L. Maier-Hein, M. Eisenmann, A. Reinke, S. Onogur, M. Stankovic, P. Scholz, T. Arbel, H. Bogunovic, A. P. Bradley, A. Carass, et al., Why rankings of biomedical im- age analysis competitions should be interpreted with care, Nature communications 9 (2018) 5217

2018
[30]

DINOv3

O. Siméoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, et al., Dinov3, arXiv preprint arXiv:2508.10104 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[31]

S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, S. Xie, Convnext v2: Co-designing and scaling con- vnets with masked autoencoders, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 16133–16142

2023
[32]

Caron, H

M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, A. Joulin, Emerging properties in self- supervised vision transformers, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

2021
[33]

J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, T. Kong, ibot: Image bert pre-training with online tok- enizer, arXiv preprint arXiv:2111.07832 (2021)

work page internal anchor Pith review arXiv 2021
[34]

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, et al., LoRA: Low-Rank Adaptation of Large Language Models., Iclr 1 (2022) 3

2022
[35]

Billot, D

B. Billot, D. N. Greve, O. Puonti, A. Thielscher, K. Van Leemput, B. Fischl, A. V . Dalca, J. E. Iglesias, et al., SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining, Medical image analysis 86 (2023) 102789

2023
[36]

Fischl, FreeSurfer, Neuroimage 62 (2012) 774–781

B. Fischl, FreeSurfer, Neuroimage 62 (2012) 774–781

2012
[37]

LaBella, O

D. LaBella, O. Khanna, S. McBurney-Lin, R. Mclean, P. Nedelec, A. S. Rashid, N. H. Tahon, T. Altes, U. Baid, R. Bhalerao, et al., A multi-institutional meningioma MRI dataset for automated multi-sequence image segmenta- tion, Scientific data 11 (2024) 496

2024
[38]

D. P. Kingma, M. Welling, Auto-Encoding Variational Bayes, arXiv preprint arXiv:1312.6114 (2013)

work page internal anchor Pith review Pith/arXiv arXiv 2013
[39]

P. M. Gordaliza, J. Banus, B. Gérin, M. Wynen, N. Molchanova, J. Richiardi, M. B. Cuadra, From 100,000+images to winning the first brain mri foun- dation model challenges: Sharing lessons and models (2026). URL:https://arxiv.org/abs/2601.13166. arXiv:2601.13166

work page arXiv 2026
[40]

Decoupled Weight Decay Regularization

I. Loshchilov, F. Hutter, Decoupled Weight Decay Regu- larization, arXiv preprint arXiv:1711.05101 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[41]

M. Beck, K. Pöppel, M. Spanring, A. Auer, O. Prud- nikova, M. Kopp, G. Klambauer, J. Brandstetter, S. Hochreiter, xLSTM: Extended Long Short-Term Mem- ory, Advances in Neural Information Processing Systems 37 (2024) 107547–107603

2024
[42]

D. P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization, arXiv preprint arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014
[43]

Hatamizadeh, V

A. Hatamizadeh, V . Nath, Y . Tang, D. Yang, H. R. Roth, D. Xu, Swin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images, in: Inter- national MICCAI brainlesion workshop, Springer, 2021, pp. 272–284

2021
[44]

X. Chen, H. Fan, R. Girshick, K. He, Improved base- lines with momentum contrastive learning, arXiv preprint arXiv:2003.04297 (2020)

work page internal anchor Pith review arXiv 2003
[45]

Y . He, V . Nath, D. Yang, Y . Tang, A. Myronenko, D. Xu, Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d medical image segmentation, in: In- ternational Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, 2023, pp. 416– 426

2023
[46]

Z. Xie, Z. Zhang, Y . Cao, Y . Lin, J. Bao, Z. Yao, Q. Dai, H. Hu, SimMIM: A Simple Framework for Masked Im- age Modeling, in: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, 2022, pp. 9653–9663

2022
[47]

Vaish, F

P. Vaish, F. Meister, T. Heimann, C. Brune, J. M. Wolterink, Consistent View Alignment Improves Founda- tion Models for 3D Medical Image Segmentation, arXiv preprint arXiv:2509.13846 (2025)

work page arXiv 2025
[48]

La Rosa, J

F. La Rosa, J. Dos Santos Silva, E. Dereskewicz, A. In- vernizzi, N. Cahan, J. Galasso, N. Garcia, R. Graney, S. Levy, G. Verma, et al., BrainAgeNeXt: advancing 15 brain age modeling for individuals with multiple sclero- sis, Imaging Neuroscience 3 (2025) imag_a_00487

2025
[49]

S. Roy, G. Koehler, C. Ulrich, M. Baumgartner, J. Pe- tersen, F. Isensee, P. F. Jaeger, K. H. Maier-Hein, Med- next: transformer-driven scaling of convnets for medical image segmentation, in: International conference on med- ical image computing and computer-assisted intervention, Springer, 2023, pp. 405–415

2023
[50]

STU-Net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training

Z. Huang, H. Wang, Z. Deng, J. Ye, Y . Su, H. Sun, J. He, Y . Gu, L. Gu, S. Zhang, Y . Qiao, STU-Net: Scalable and Transferable Medical Image Segmentation Models Em- powered by Large-Scale Supervised Pre-training, arXiv preprint arXiv:2304.06716 (2023)

work page arXiv 2023
[51]

Wasserthal, H.-C

J. Wasserthal, H.-C. Breit, M. T. Meyer, M. Pradella, D. Hinck, A. W. Sauter, T. Heye, D. T. Boll, J. Cyriac, S. Yang, et al., TotalSegmentator: robust segmentation of 104 anatomic structures in CT images, Radiology: Artifi- cial Intelligence 5 (2023) e230024

2023
[52]

Dancette, J

C. Dancette, J. Khlaut, A. Saporta, H. Philippe, E. Fer- reres, B. Callard, T. Danielou, L. Alberge, L. Machado, D. Tordjman, et al., Curia: A multi-modal foundation model for radiology, arXiv preprint arXiv:2509.06830 (2025)

work page arXiv 2025
[53]

K. Li, Y . Wang, J. Zhang, P. Gao, G. Song, Y . Liu, H. Li, Y . Qiao, Uniformer: Unifying convolution and self-attention for visual recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (2023) 12581–12600

2023
[54]

S. Rui, L. Chen, Z. Tang, L. Wang, M. Liu, S. Zhang, X. Wang, Multi-modal vision pre-training for medical im- age analysis, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5164–5174. Appendix A. Dataset Details This appendix details subject demographics, MRI sequences, preprocessing, and labeling protocols for all FOMO...

2025