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

Abstract

We present FOMO260K, a large-scale, heterogeneous dataset of 260,927 brain Magnetic Resonance Imaging (MRI) scans from 77,589 MRI sessions and 55,378 subjects, aggregated from 910 publicly available sources. The dataset includes both clinical- and research-grade images, multiple MRI sequences, and a wide range of anatomical and pathological variability, including scans with large brain anomalies. Minimal preprocessing was applied to preserve the original image characteristics while reducing entry barriers for new users. Companion code for self-supervised pretraining and finetuning is provided, along with pretrained models. FOMO260K is intended to support the development and benchmarking of self-supervised learning methods in medical imaging at scale.

Background & Summary

Self-supervised learning (SSL) has led to major breakthroughs in computer vision and natural language processing, largely driven by the availability of large-scale public datasets such as ImageNet¹, Places365², and OpenWebText³. These resources have enabled the development, benchmarking, and rapid iteration of powerful SSL methods under standardized settings. In neuroimaging, however, the lack of comparably large and diverse public datasets has slowed the adoption and evaluation of SSL approaches. Existing large-scale datasets, such as ADNI⁴, UK Biobank⁵, PPMI⁶, and ABCD⁷, while valuable, are often curated for specific diseases or patient populations, and typically follow homogeneous imaging protocols with limited pathological variability. Access is often restricted by formal applications, strict data use agreements, and institutional approvals, and data are commonly distributed in formats that require domain-specific preprocessing (e.g., conversion from DICOM to NIfTI format or handling 4D acquisitions). These challenges raise the barrier to entry and hinder the scalability of SSL pretraining.

To address these concerns, we introduce FOMO260K, a large-scale, heterogeneous dataset of brain MRI scans, comprising 260,927 scans from 77,589 sessions and 55,378 subjects, aggregated from 910 publicly available sources. FOMO260K includes both clinical- and research-grade imaging across multiple MRI sequences and captures a wide range of anatomical and pathological variability, including scans with large brain anomalies, making it more representative of real-world population-level data. FOMO260K builds on recent large-scale aggregation efforts such as OpenMind⁸, which provides an important and well-curated resource of 114 K scans from OpenNeuro⁹. Almost all of these scans are also included in FOMO260K, which is more than two times larger and exhibits greater heterogeneity in imaging protocols and slice thickness, including a higher proportion of low-resolution scans typical of clinical practice. An overview of the differences between the two datasets is shown in Table 1.

Table 1.

Comparison between OpenMind, FOMO45K and FOMO260K.

Dataset	Scans	Sessions	Co-registered	Disease Metadata	Clinical-grade (scans)	Control	Brain Tumor	Stroke	Mental Disorders	Dementia	Neurological Disorders
OpenMind	114 K	46 K (35 K)	✗	✗	8 K (7%)	30 K (85%)	0.2 K (1%)	2.5 K (7%)	1.2 K (4%)	0 (0%)	0.7 K (2%)
FOMO45K (ours)	46 K	12 K (8 K)	✓	✓	5 K (11%)	2.7 K (34%)	2.7 K (34%)	2.5 K (30%)	0 (0%)	0.3 K (3%)	0 (0%)
FOMO260K (ours)	261 K	78 K (64 K)	✗	✓	53 K (20%)	41 K (64%)	16 K (24%)	2.6 K (4%)	2.1 K (3%)	0.3 K (1%)	0.9 K (1%)

FOMO260K encompasses a broader range of disorders and a higher proportion of low-resolution scans, which are typical in clinical practice. In particular, FOMO260K contains considerably more tumors compared to OpenMind. We use slice thickness as a proxy for classifying a scan as “clinical-grade”, and define a clinical-grade scan as having a slice thickness above 3 mm. We report disorder groups by number of sessions, derived from the disease metadata as described in Methods. Numbers in parentheses indicate the number of sessions with available disease metadata, and percentages are computed with respect to this subset. Note that nearly all scans in OpenMind and FOMO45K are included in FOMO260K.

FOMO260K is available as a single download from Hugging Face¹⁰, together with code to preprocess the scans, incorporate additional datasets, perform self-supervised pretraining, and finetune models. This setup enables method development, standardized benchmarking, reproducible experiments, and broader adoption of SSL in medical imaging. An earlier release of this dataset, FOMO45K, was developed in parallel with the Foundation Model challenge at MICCAI 2025 (FOMO25)¹¹, which aims to catalyze progress in self-supervised learning for medical imaging. Since then, the dataset has been significantly expanded to form FOMO260K, and it will continue to grow in future releases as additional cohorts and modalities become available.

Methods

Input data sources and acquisition

FOMO260K contains 260,927 scans from 77,589 MRI sessions and 55,378 subjects, aggregated from 910 publicly available sources. All available scans from each source were eligible for inclusion, except ex vivo scans, functional MRI, field maps, SWI phase images, positron emission tomography scans, and computed tomography scans, which we excluded to prevent distribution shifts that could hinder SSL pretraining. No other exclusion criteria were applied. Table 2 summarizes the source datasets, listing the number of subjects, sessions, scans, MRI sequence types, and license. All included datasets are publicly available under licenses compatible with CC BY-NC-SA redistribution, and the corresponding citations provide the DOI, URL, or accession identifier required to retrieve each original dataset. The largest contributions of FOMO260K come from HBN, Yale Brain Mets Longitudinal, and the OpenNeuro repository. T1-weighted structural MRI is present in 876 of the 910 constituent datasets (96.3%), followed by T2-weighted (22.6%), diffusion-weighted (13.2%), FLAIR (3.5%), and PD (2.2%) sequences. A visual overview of the dataset’s heterogeneity–in terms of image quality, modality, and pathology–is provided in Fig. 1.

Table 2.

Overview of the current datasets in FOMO260K.

Source dataset	Subjects	MRI sessions	MRI scans	MRI sequences	License
ClevelandCCF39,51	31	31	31	T1w	CC BY-NC
Nigerian Clinical40,65	82	140	701	T1w, T2w, FLAIR, DWI	CC BY 4.0
CUNMET41	51	59	173	T1w, M0, ASL	CC BY-NC
ACPI42,51	163	163	163	T1w	CC BY-NC
ADHD_20043,51	973	973	973	T1w	CC BY-NC
AHEAD44,66	105	105	420	T1w, T1map, R1map, R2*map	CC BY 4.0
ATAG45,67	53	53	583	T2*w, MP2RAGE	CC0
Adolescent Brain Development46,68	51	96	905	T1w, T2w, DWI, ASL, CBF	PDDL
BraTS-GEN28–32,69	2,611	2,611	10,448	T1w, T2w, FLAIR, T1ce	CC BY-NC
BrainLat70,71	592	592	2,211	T1w, FLAIR, DWI	CC0
CFMM-7T47,72	32	32	160	T1w, MP2RAGE	CC BY 4.0
CHBMP48,73	203	203	901	T1w, DWI	CC BY-NC-SA
Calgary Preschool49,74	162	473	1,629	T1w, DWI, CBF	CC BY 4.0
CoRR50	1,498	2,779	5,831	T1w, MP2RAGE, DWI, ASL, CBF	Open Access
MSD Brain Tumor33–36,75	750	750	3,000	T1w, T2w, FLAIR, T1ce	CC BY-SA 4.0
HBN-SSI51,52	13	168	1,031	T1w, T2w, DWI	CC BY-NC-SA 4.0
HBN53,76	3,847	3,849	26,097	T1w, T2w, PD, FLAIR, DWI	CC BY-NC-SA 4.0
IXI37	584	584	3,105	T1w, T2w, PD, DWI, ANGIO	CC BY-SA 3.0
Beijing Enhanced51,54	180	180	540	T1w, DWI	CC BY-NC
Infant Development Brain55,77	833	833	1,666	T1w, T2w	CC BY 4.0
M4Raw56,78	208	208	2,230	T1w, T2w, FLAIR	CC BY 4.0
MICA MICs79,80	50	50	649	T1w, T1map, MP2RAGE	CC0
NKI38,81	1,327	2,456	10,658	T1w, T2w, DWI	Open Access
OpenNeuro (884 datasets)82–959	37,949	45,377	140,389	T1w, MP2RAGE, T2w, T2w*, T1ce, FLAIR, DWI, T1map, T2map, T2*map, GRE, ANGIO, CBF, ASL, M0, minIP, SWI	CC0
SLIM51,57	594	1,048	3,096	T1w, DWI	CC BY-NC
Tao Wu51,58	40	40	40	T1w	CC BY-NC-SA
WAND59,960	177	850	7,132	T1w, MP2RAGE, T2w, DWI, ANGIO, CBF, ASL, M0	CC BY 4.0
Wayne51,60	426	610	610	T1w	CC BY-NC-SA
Yale Brain Mets Longitudinal61,62	1,430	11,877	33,800	T1w, T2w, FLAIR, T1ce	CC BY 4.0
Yale High Res63	232	268	576	T1w	CC BY-NC-SA
Age ility64,961	131	131	1,179	T1w, DWI	CC BY-NC-SA
FOMO260K	55,378	77,589	260,927		CC BY-NC-SA 4.0

For each dataset, we summarize the number of subjects, MRI sessions, scans, available sequence types, and license. Sequences are abbreviated as follows: T1 = T1-weighted, T2 = T2-weighted, T2* = T2*-weighted, T1ce = T1-weighted contrast-enhanced, FLAIR = Fluid-Attenuated Inversion Recovery, DWI = Diffusion-Weighted Imaging, PD = Proton Density, SWI = Susceptibility Weighted Imaging, GRE = Gradient Echo, minIP = minimum intensity projection, MP2RAGE = Magnetization-Prepared 2 Rapid Acquisition Gradient Echoes, ANGIO = Angiography, ASL = Arterial Spin Labeling, M0 = Proton Density reference image for ASL, CBF = Cerebral Blood Flow, T1map = T1 relaxation time map, T2map = T2 relaxation time map, R1map = R1 (1/T1) relaxation rate map, R2map = R2 relaxation rate map.

Fig. 1.

Representative examples from the FOMO260K dataset, illustrating the heterogeneity in image quality, MRI sequences, and the presence of brain anomalies.

MRI preprocessing

We minimally preprocessed all the MRI scans to ensure consistent brain orientation and comparable input dimensionality across datasets. To achieve this, all scans were reoriented to RAS (Right-Anterior-Superior) orientation, 4D volumes were converted to 3D, and scans with fewer than 15 slices were discarded due to insufficient anatomical coverage. For DWI comprising several frames corresponding to diffusion-weighted images with different gradient directions, each volume’s b-value and gradient vector were used for selection. If b0 (b = 0) or a near-b0 (b ≤ 5) was available, it was extracted and saved separately as a reference to provide a standard non-diffusion-weighted reference. The remaining diffusion-weighted volumes were grouped into b-value shells (allowing up to 50 units of difference), and for each shell, three volumes whose gradient directions were closest to the canonical x, y, and z axes were selected (measured in terms of cosine similarity). To increase variability in DWI representations during SSL pretraining, two preprocessing strategies were applied: for half of the scans, these three volumes were averaged to produce a single 3D image representing the diffusion contrast of the shell; for the other half, a trace apparent diffusion coefficient (ADC) image was computed using the three axis-aligned volumes and the b0 reference according to:

$$\begin{array}{cc}\mathrm{Trace}={\mathrm{ADC}}_x+{\mathrm{ADC}}_y+{\mathrm{ADC}}_z,& \mathrm{where}{\mathrm{ADC}}_i=-\frac{1}{b_i}\ln \frac{S_i}{S_0}\end{array}$$

and where $S_i$ is the signal intensity of the diffusion-weighted volume along axis i, $S_0$ is the signal intensity of the corresponding b0 or near-b0 reference, and $b_i$ is the b-value corresponding to the i-th axis-aligned volume. To ensure numerical stability and exclude noise artifacts, ADC was computed only for voxels where $S_0>0$, $S_i>0$, and $S_i\le S_0$, with the signal ratio clipped to ${\mathrm{[10}}^{-10}\mathrm{,1.0]}$ before taking the logarithm; voxels failing these criteria were set to zero for that direction. If no b0 or near-b0 reference was available, the three axis-aligned volumes were averaged instead. For perfusion imaging, preprocessing depended on the number of channels. When more than three channels (e.g., multiple M0 volumes) were present, all channels were averaged. For scans with two or three channels, such as ASL acquisitions, the perfusion-weighted volume was obtained by subtracting the first channel from the last.

Demographic and MRI information collection

Demographic information, including age, sex, handedness, and diagnostic group, as well as MRI acquisition details, was collected for all scans. Subject-level information was extracted from dataset-provided files or, when unavailable, from the original publications or general dataset descriptions. Each MRI session was assigned to the “control” groups when they corresponded to healthy individuals or patients without any major medical, neurological, or psychiatric conditions. For visualization and summary statistics, related diagnostic groups were merged into broader categories (Controls, Brain Tumor, Mental Disorders, Dementia, Neurological Disorders, Memory Complaints, Stroke, and Other) using keyword-based matching. The complete group mapping algorithm is available in the publicly available data processing code repository. When information could not be determined with confidence, it was omitted. MRI acquisition metadata were obtained either from files accompanying the scans or from the original publications, including general scanner and sequence information.

Dataset composition and demographics

Figure 2 summarizes the overall composition of the dataset. A total of 57,224 MRI sessions have available age information at each MRI session, spanning the entire lifespan with a distribution peaking in young adulthood (mean age 34.8 years). The sex distribution is relatively balanced, with a slight predominance of females (53%). Handedness data are available for 13,768 MRI sessions, with the majority of the subjects being right-handed (91%). The dataset includes a broad range of subject groups. Controls account for 64% of the MRI sessions with available diagnostic information. The remaining sessions comprise individuals with brain tumors (24%), stroke (4%), mental disorders (3%), neurological disorders (1%), and other clinical conditions (2%). MRI data were acquired across multiple scanner manufacturers, predominantly Siemens (80%) and Philips (10%). Field strengths are primarily 3 T (81%) and 1.5 T (15%), with both low-field and ultra-high-field scans also represented. The dataset includes both 2D (54%) and 3D (46%) acquisitions. The most common sequence types are diffusion-weighted imaging (31%), T1-weighted (31%), and T2-weighted (12%). Slice thickness varies across datasets, reflecting the heterogeneous acquisition protocols used across sites.

Fig. 2.

(A) Age distribution of participants at each MRI session. (B) Distribution of subject groups at each MRI session. (C) Sex distribution at each MRI session (F = female, M = male). (D) Handedness distribution at each MRI session (R = right-handed, L = left-handed, A = ambidextrous). (E) Distribution of MRI scanner manufacturers across all scans. (F) Acquisition types showing the proportion of 2D and 3D sequences. (G) Field strength distribution. (H) Top 15 scanner models, with the MRI scanner manufacturer indicated by color coding. (I) Slice thickness distribution across the dataset (note: may include thickness of already resampled data from source datasets). (J) Top 15 MRI sequences, with modality indicated by color coding. Percentages are reported relative to the subset of data entries with available information for each variable.

FOMO45K

FOMO45K contains 46,149 MRI scans from 11,967 MRI sessions across 9,490 subjects, aggregated from 13 publicly available datasets. The scans in FOMO45K correspond to a subset of those included in FOMO260K, but the two releases differ in their preprocessing. They are intended to serve complementary purposes: FOMO260K provides minimally preprocessed scans that preserve the original image characteristics, whereas FOMO45K provides co-registered and either skull-stripped or defaced versions of the same scans, ready for downstream tasks that require standardized, anatomically aligned inputs (see preprocessing details below). Table 3 summarizes the source datasets, including the number of subjects, sessions, and scans; MRI sequence types; preprocessing steps applied; and licenses. All included datasets are publicly available under licenses compatible with CC BY-NC-SA redistribution, and the corresponding citations provide the DOI, URL, or accession identifier required to retrieve each original dataset.

Table 3.

Overview of the datasets in FOMO45K.

Source dataset	Subjects	MRI sessions	MRI scans	MRI sequences	Skull-stripped/Defaced	License
OpenNeuro - SOOP755	1,715	1,715	6,508	T1, FLAIR, DWI	Defaced	CC0
BraTS-GEN28–32,69	2,251	2,251	9,004	T1, T2, FLAIR, T1ce	Skull-stripped	CC BY-NC
MSD Brain Tumor33–36,75	484	484	1,936	T1, T2, FLAIR, T1ce	Skull-stripped	CC BY-SA 4.0
IXI37	584	584	2,530	T1, T2, PD, DWI	Skull-stripped	CC BY-SA 3.0
OpenNeuro - NIMH899	249	252	2,089	T1, T2, T2*, DWI	Skull-stripped	CC0
OpenNeuro - DLBS749	464	957	3,845	T1, T2, DWI	Skull-stripped	CC0
OpenNeuro - IDEAS882	542	542	1,035	T1, FLAIR	Defaced	CC0
OpenNeuro - ARC754	230	741	2,029	T1, T2, DWI	Defaced	CC0
OpenNeuro - MBSR775	147	348	1,023	T1, DWI	Defaced	CC0
OpenNeuro - UCLA92	265	265	789	T1, DWI	Defaced	CC0
OpenNeuro - QTAB616	417	721	6,760	minIP, GRE, SWI	Defaced	CC0
NKI38,81	1,316	2,281	4,502	T1, T2, DWI	Defaced	Open Access
OpenNeuro - AOMIC ID1000432	826	826	4,099	T1, DWI	Defaced	CC0
FOMO45K	9,490	11,967	46,149			CC BY-NC-SA 4.0

For each dataset, we summarize the number of subjects, MRI sessions, scans, available sequence types, whether the images were skull-stripped or defaced, and license. See Methods for full preprocessing details. Sequences are abbreviated as follows: T1 = T1-weighted, T2 = T2-weighted, T2* = T2*-weighted, T1ce = T1-weighted contrast-enhanced, FLAIR = Fluid-Attenuated Inversion Recovery, DWI = Diffusion-Weighted Imaging, PD = Proton Density, SWI = Susceptibility Weighted Imaging, GRE = Gradient Echo, minIP = minimum intensity projection.

MRI preprocessing for FOMO45K comprised three main steps: reorienting images to RAS orientation (as in FOMO260K), affine co-registration, and skull-stripping. All scans were first reoriented to RAS and affinely co-registered using the mri_coreg command from FreeSurfer 7.4.1¹², with default parameters. Within each MRI session, scans were aligned to the image with the highest spatial resolution to preserve most anatomical detail.

For DWI scans in 4D format, if a b0 was available, it was extracted and saved separately. For the b1000 shell, the volumes were processed as in FOMO260K and averaged into a single representative 3D image. In contrast to FOMO260K, no trace ADC computation or additional b-value shell grouping was performed.

Skull-stripping was performed using SynthSeg¹³ (FreeSurfer 7.4.1), which outputs segmentation masks of brain structures. These masks were used to define the brain extraction region. Skull-stripping was only applied when the images were not already defaced or skull-stripped by the dataset provider.

Ethics statement

All data included in FOMO260K and FOMO45K were obtained from publicly available sources under open licenses; the original license of each constituent dataset is listed in Tables 2 and 3. Each constituent dataset was collected under protocols approved by the respective institutional review boards of the original studies. No new human subjects research was conducted for this data aggregation project.

Data Record

FOMO260K¹⁰ is publicly available through https://doi.org/10.57967/hf/8670, while FOMO45K¹⁴ is publicly available at https://doi.org/10.57967/hf/8669. All MRI scans are stored in NIfTI-compressed format and organized using a modified version of the Brain Imaging Data Structure (BIDS) format¹⁵. For most modalities, the directory structure and naming closely follow BIDS conventions. However, several deviations were introduced to ensure scalability and consistency across FOMO260K. First, acquisition and scanner metadata are not stored as individual JSON files per scan; instead, MRI acquisition information is centralized in a single tabular file (mri_info.tsv). Second, when sequence information was unavailable or ambiguous, scan names were kept close to their original labels or generically named scan. Third, diffusion scans do not follow the BIDS DWI specification: they are provided as derived 3D volumes without accompanying bval and bvec files, with the b-value group and whether a trace ADC representation (_trace) was used explicitly encoded in the filename. Finally, subject-level information and scan provenance are provided via tabular files: participants.tsv contains demographic information and group assignments, while mapping.tsv records the correspondence between each scan and its original source dataset. Full details are documented in the accompanying code repository.

Technical Validation

Empirical validation

To validate the efficacy of the FOMO260K dataset, we pretrain a model on FOMO260K using a self-supervised masked autoencoder objective, in which the model learns general-purpose representations by reconstructing masked portions of the input without relying on any task-specific labels. We then show that the pretrained model achieves better downstream performance than a model trained from randomly initialized weights (referred to as “scratch” in the remainder of the paper). We evaluate the result of pretraining in the challenging few-shot setting, where the model must generalize with limited supervision, making any improvement a direct indicator that the pretrained features transfer effectively to new tasks.

Pretraining configuration

We train a 100M-parameter ResEnc¹⁶ model using AMAES (Augmented Masked Auto Encoder for 3D Segmentation)¹⁷, a variant of the popular masked-autoencoder pretraining strategy optimized for representation learning on 3D MRI data¹⁸, using a masking ratio of 60% and a masking unit size of $8\times 8\times 8$ voxels. Training used eight H100 GPUs for 44 hours with a global batch size of 64 for 187,500 steps. Each scan was processed using a patch size of $160\times 160\times 160$. We optimize with AdamW¹⁹ at a base learning rate of $1\times {10}^{-4}$ under a cosine decay schedule, with 2% linear warmup and gradient accumulation of two steps.

Finetuning configuration

We evaluate models on a diverse set of segmentation tasks using the ISLES22²⁰ ($n\mathrm{=250}$), ATLAS²¹ ($n\mathrm{=655}$), SBM3^22,23 ($n\mathrm{=105}$), WMH²⁴ ($n\mathrm{=170}$), and Cerebrum-7T²⁵ ($n\mathrm{=142}$) datasets. All datasets are used in a few-shot setting, using 20 labeled examples per training split. Models are trained for 37,500 steps with AdamW, a max learning rate of $1\times {10}^{-3}$. Only the encoder weights are transferred. To adapt the decoder to the encoder, we first freeze the encoder for 2500 steps, training only the decoder with a linearly increasing learning rate. The encoder and decoder are then linearly warmed up for 2,500 steps before the entire network is trained with a cosine decay schedule. We use a batch size of two. Scratch baselines follow an nnU-Net-inspired configuration²⁶, employing stochastic gradient descent with a base learning rate of $1\times {10}^{-2}$ using a cosine decay schedule for 37,500 steps.

Validation results

Results are provided in Table 4. We report the average Dice score over five folds along with the standard error of the mean. If the dataset contained multiple classes, we report the average of the foreground classes. The pretrained model consistently outperforms the scratch baseline across all five benchmark datasets, with the differences statistically significant at the $p\mathrm{<0.05}$ level (two-sided paired t-test on per-case Dice differences, Holm-corrected²⁷ across datasets). The largest gains are observed on SBM3 (+1.45 Dice), WMH (+1.29 Dice), and ISLES22 (+1.12 Dice), with smaller but still consistent gains on ATLAS, and Cerebrum-7T.

Table 4.

Dice performance (mean ± standard error of the mean) across five-fold cross-validation.

Method	ATLAS	ISLES22	SBM3	WMH	Cerebrum-7T
Scratch	$41.49\pm 0.72$	$72.86\pm 0.89$	$58.23\pm 1.44$	$72.75\pm 0.58$	$90.54\pm 0.12$
AMAES on FOMO260K	$\mathbf{42}\mathbf{.}\mathbf{43}\pm \mathbf{1}\mathbf{.}\mathbf{36}$	$\mathbf{73}\mathbf{.}\mathbf{98}\pm \mathbf{0}\mathbf{.}\mathbf{64}$	$\mathbf{59}\mathbf{.}\mathbf{68}\pm \mathbf{1}\mathbf{.}\mathbf{03}$	$\mathbf{74}\mathbf{.}\mathbf{04}\pm \mathbf{0}\mathbf{.}\mathbf{51}$	$\mathbf{90}\mathbf{.}\mathbf{80}\pm \mathbf{0}\mathbf{.}\mathbf{06}$

Pretraining on FOMO260K yields consistently higher accuracy than training from scratch. Bold values indicate $p\mathrm{<0.05}$ for a two-sided paired t-test on per-case Dice differences, with Holm correction²⁷ across datasets.

Usage Notes

FOMO260K¹⁰ and FOMO45K¹⁴ are distributed under the CC BY-NC-SA 4.0 license. The original licensing terms of each constituent dataset are listed in Table 2 for FOMO260K and Table 3 for FOMO45K. To ensure proper attribution and recognition of the source datasets, all users must cite the following papers for FOMO45K and FOMO260K: BraTS-GEN^28–32, MSD Brain Tumor^33–36, IXI³⁷, NKI³⁸.

Additionally, the following papers must be cited for FOMO260K: ClevelandCCF³⁹, Nigerian Clinical⁴⁰, CUNMET⁴¹, ACPI⁴², ADHD_200⁴³, AHEAD⁴⁴, ATAG⁴⁵, Adolescent Brain Development⁴⁶, CFMM-7T⁴⁷, CHBMP⁴⁸, Calgary Preschool⁴⁹, CoRR⁵⁰, HBN-SSI^51,52, HBN⁵³, Beijing Enhanced^51,54, Infant Development Brain⁵⁵, M4Raw⁵⁶, SLIM^51,57, Tao Wu^51,58, WAND⁵⁹, Wayne^51,60, Yale Brain Mets Longitudinal^61,62, Yale High Res⁶³, and Age ility⁶⁴.

Author contributions

S.C. and A.M. contributed equally to this work and were responsible for conceptualization, methodology, data curation, software development, formal analysis, validation, visualization, and writing of the original draft. S.N.L. contributed to methodology, software development, data curation, and manuscript review and editing. J.A. contributed to methodology, software development, formal analysis, and manuscript review and editing. J.M. contributed to software development, data curation, and manuscript review and editing. P.R.G. contributed to software development and manuscript review and editing. V.N., C.H.K., P.L., M.M.G., M.B., and M.E.B. contributed to manuscript review and editing. J.E.I. contributed to supervision, resources, and manuscript review and editing. M.N. contributed to supervision and manuscript review and editing. All authors reviewed and approved the final manuscript.

Funding

This work has been supported by the Danish Data Science Academy, which is funded by the Novo Nordisk Foundation (grant number NNF21SA0069429) and Villum Fonden (grant number 40516), Pioneer Centre for AI, Danish National Research Foundation, grant number P1, the Lundbeck Foundation (grant number R449-2023-1512), the Novo Nordisk Foundation (grant number 0104988 for the Gefion AI Supercomputer), and the National Institute of Health (grant number 1R01AG070988, 1RF1AG080371, 1RF1MH123195, 1UM1MH130981, 1R21NS138995, and 1R01EB031114).

Data availability

FOMO260K¹⁰ and FOMO45K¹⁴ are publicly available on Hugging Face at https://doi.org/10.57967/hf/8670 and https://doi.org/10.57967/hf/8669, respectively, under the CC BY-NC-SA 4.0 license.

Code availability

The preprocessing scripts for FOMO260K are publicly available at https://github.com/Sllambias/asparagus_preprocessing. The code used for pretraining the models is available at https://github.com/Sllambias/asparagus. The preprocessing scripts for FOMO45K, together with additional code required to reproduce all analyses, tables, and figures reported in this manuscript, are available at https://github.com/FGA-DIKU/fomo_mri_datasets. Pretrained model weights are publicly released at https://huggingface.co/FOMO-MRI/AMAES_resenc_b_fomo260k.

Competing interests

M.N. holds shares in Cerebriu.