FOMO260K
FOMO260K brain MRI dataset for large-scale self-supervised learning
Openly downloadable aggregate of 260,927 minimally preprocessed brain MRI scans from 77,589 sessions and 55,378 subjects, pooled from 910 public sources, with clinical- and research-grade scans of many sequences for self-supervised pretraining.
Overview
FOMO260K is a pooled collection of brain MRI built for self-supervised pretraining of 3D models. A team at the University of Copenhagen and the Pioneer Centre for AI gathered scans from 910 public sources, applied only light preprocessing and republished everything as one Hugging Face download that needs no account. Companion code covers preprocessing, pretraining and finetuning, and pretrained weights are released with it.
Composition
The release holds 260,927 scans from 77,589 sessions of 55,378 people. Most scans come from HBN, the Yale Brain Mets Longitudinal cohort and 884 OpenNeuro datasets, which alone contribute 140,389 scans. Sequences include T1w, T2w, FLAIR, contrast-enhanced T1w, PD, T2*w, SWI magnitude, minIP, diffusion (derived 3D volumes and trace ADC), MP2RAGE, relaxometry maps, angiography and ASL. Of the sessions with diagnostic metadata, about two thirds are controls and about a quarter have brain tumors; stroke, mental disorders, neurological disorders and dementia make up small shares. Ages span the lifespan, with a session-level mean of 34.8 years.
Acquisition
Scans come from many sites and protocols. The paper reports mostly Siemens (80%) and Philips (10%) scanners, mostly 3T
(81%) and 1.5T (15%) field strength, and roughly half 2D acquisitions. About 20% of scans have slices thicker than 3 mm,
which the authors treat as clinical grade. Images were reoriented to RAS, 4D series were reduced to 3D and scans with
fewer than 15 slices were dropped. Functional MRI, field maps, SWI phase, PET, CT and ex vivo scans were excluded.
Files follow a modified BIDS layout with participants.tsv, mapping.tsv (link to the source scan) and
mri_info.tsv (acquisition parameters).
Annotations
No image labels. Age, sex, handedness and a diagnostic group are given where the source provided them.
Related releases
FOMO-MRI also publishes FOMO300K (306,207 scans in V1.1 of April 2026), the full superset that adds sources under data use agreements (OASIS, GSP, HCP WU-Minn) and is gated with automatic approval; FOMO50K (49,193 scans), a co-registered and skull-stripped or defaced subset first released as FOMO60K for the FOMO25 challenge; and FOMO45K (46,149 scans), the open part of FOMO50K. Each is a subset of FOMO300K, so they must not be combined.
Known limitations
- Demographic and diagnostic metadata are incomplete and were harmonised by keyword matching.
- Scanner metadata in
mri_info.tsvare not normalised across sources. - Constituent datasets keep their own licenses and citation duties.
Cohort
Aggregate numbers from the sources below. Bars are relative to the 55,378 subjects.
Sex
- Female 20,510 37%
- Male 18,704 34%
- Unknown 16,164 29%
Age
mean 34.8No age bins reported.
Condition
studies, values can overlap
- Healthy control ~41,000 53%
- Brain tumor ~16,000 21%
- Stroke ~2,600 3%
- Dementia ~300 <1%
License and access
Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer
Download without an account
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
Non-commercial use only, with credit. Anything you share that is built on the data must use the same license.
The constituent datasets keep their own licenses (CC0, PDDL, CC BY, CC BY-SA, CC BY-NC, CC BY-NC-SA and "Open Access", listed in Table 2 of the paper), and users must also cite the constituent datasets named in the paper's Usage Notes.
What you can do
- No
- Conditional
- Yes
- Yes
What you can share
- Conditional
- Conditional
- Conditional
What you must do
- Yes
- Yes
- No
- No
- Manuscript review No
- Release code No
- Return results No
- Delete after use No
Limits
- No
- Location limits No
Citation
Cerri S, Munk A, Llambias SN, Ambsdorf J, Machnio J, Nersesjan V, Hedeager Krag C, Liu P, Rocamora García P, Mehdipour Ghazi M, Boesen M, Benros ME, Iglesias JE, Nielsen M. A large-scale heterogeneous 3D magnetic resonance brain imaging dataset for self-supervised learning. arXiv:2506.14432 (2026).
@article{Cerri2026large,
title={A large-scale heterogeneous 3D magnetic resonance brain imaging dataset for self-supervised learning},
author={Cerri, Stefano and Munk, Asbj{\o}rn and Llambias, Sebastian N{\o}rgaard and Ambsdorf, Jakob and Machnio, Julia and Nersesjan, Vardan and Hedeager Krag, Christian and Liu, Peirong and Rocamora Garc{\'\i}a, Pablo and Mehdipour Ghazi, Mostafa and Boesen, Mikael and Benros, Michael Eriksen and Iglesias, Juan Eugenio and Nielsen, Mads},
journal={arXiv preprint arXiv:2506.14432},
year={2026},
url={https://arxiv.org/abs/2506.14432}
} All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Subjects | total | 55,378 | cerri2026 Table 2 |
| Studies | total MRI sessions | 77,589 | cerri2026 Table 2 |
| Scans | total | 260,927 | cerri2026 Table 2 |
| Subjects | sex=female subjects with F in at least one session; 7 of them have other sessions without sex | 20,510 | fomo260k-participants columns dataset, participant_id, sex |
| Subjects | sex=male subjects with M in at least one session; 5 of them have other sessions without sex | 18,704 | fomo260k-participants columns dataset, participant_id, sex |
| Subjects | sex=unknown no sex recorded in any session | 16,164 | fomo260k-participants columns dataset, participant_id, sex |
| Studies | condition=healthy "41K" control sessions, of 64K sessions with disease metadata | ~41,000 | cerri2026 Table 1 |
| Studies | condition=brain_tumor "16K" sessions | ~16,000 | cerri2026 Table 1 |
| Studies | condition=stroke "2.6K" sessions | ~2,600 | cerri2026 Table 1 |
| Studies | condition=dementia "0.3K" sessions | ~300 | cerri2026 Table 1 |
| Mean age | total age at each MRI session for the 57224 sessions with age information | 34.8 | cerri2026 Dataset Composition and Demographics |
Sources
The keys used in the table above.
- cerri2026 Cerri et al., A large-scale heterogeneous 3D MRI brain imaging dataset for self-supervised learning, arXiv:2506.14432v3 paper
- fomo260k-card FOMO260K dataset card on Hugging Face website
- fomo260k-participants FOMO260K participants.tsv (CC BY-NC-SA 4.0, open download), one row per session, counted per unique dataset and participant_id computed