FOMO60K: Heterogeneous 3D Brain MRI Dataset

• FOMO60K is a large-scale, heterogeneous 3D brain MRI dataset comprising 60,529 scans from 11,187 subjects, assembled for self-supervised learning and neuroimaging benchmarking.
• The dataset preserves rich scanner, site, and protocol variations by employing minimal preprocessing, thus encouraging models to learn robust invariances under clinical domain shifts.
• FOMO60K underpins the FOMO25 challenge and supports advanced tasks like MRI super-resolution, highlighting its applicability across diverse neuroimaging methods.

Searching arXiv for the most relevant FOMO60K papers and closely related usage papers. FOMO60K is a large-scale, heterogeneous 3D brain magnetic resonance imaging dataset assembled to support self-supervised learning, foundation-model pretraining, and downstream neuroimaging method development at scale. It comprises 60,529 MRI scans from 13,900 sessions across 11,187 subjects, aggregated from 16 publicly available sources, and includes both clinical- and research-grade data, multiple MRI sequence types, and substantial anatomical and pathological variability, including scans with large brain anomalies (Munk et al., 17 Jun 2025). Subsequent work positioned FOMO60K as the pretraining substrate for the FOMO25 challenge on clinical brain MRI foundation models and also used its NKI cohort as the sole imaging source for diffusion-based brain MRI super-resolution experiments (Munk et al., 13 Apr 2026, Chiche et al., 15 Mar 2026).

1. Definition and intended scope

FOMO60K was introduced as a large-scale, heterogeneous dataset of 3D brain MRI designed specifically for self-supervised learning (SSL) and benchmarking in neuroimaging. Its stated motivation is to enable ImageNet-scale-style pretraining for volumetric medical imaging while preserving the heterogeneity of real clinical and multi-cohort data. The dataset is intended to lower barriers to entry through minimal preprocessing, standardized NIfTI release, and accompanying code for pretraining and finetuning (Munk et al., 17 Jun 2025).

In the FOMO25 challenge, FOMO60K served as the central unlabeled pretraining corpus for brain MRI foundation models. The challenge explicitly used it to study whether SSL on a large, heterogeneous corpus improves robustness under domain shift and in few-shot downstream regimes. In the method track, participants were required to pretrain exclusively on FOMO60K so that differences in downstream performance could be attributed to architectural and objective choices rather than differences in pretraining data (Munk et al., 13 Apr 2026).

A common misconception is to treat FOMO60K as a conventional curated benchmark with heavy harmonization and narrow modality scope. The published description indicates the opposite: the dataset was intentionally built to retain scanner-, site-, and protocol-specific variation, and it spans multiple sequence families rather than a single structural contrast (Munk et al., 17 Jun 2025).

2. Dataset composition and heterogeneity

FOMO60K aggregates data from 16 public datasets spanning healthy cohorts, aging and neurodegeneration studies, stroke and aphasia cohorts, epilepsy surgery data, tumor datasets, and susceptibility-focused acquisitions. The included sequence types are T1, T2, FLAIR, DWI, T1ce, PD, SWI, GRE, minIP, and T2* (Munk et al., 17 Jun 2025).

Source dataset	Scans	Sequences / status
MGH Wild	11,366	Unknown; skull-stripped
SOOP	6,508	T1, FLAIR, DWI; defaced
BraTS-GEN	9,004	T1, T2, FLAIR, T1ce; skull-stripped
OASIS1	1,688	T1; skull-stripped
OASIS2	1,326	T1; skull-stripped
MSD Brain Tumor	1,936	T1, T2, FLAIR, T1ce; skull-stripped
IXI	2,530	T1, T2, PD, DWI; skull-stripped
NIMH Healthy Research Volunteer	2,089	T1, T2, T2*, DWI; skull-stripped
DLBS	3,845	T1, T2, DWI; skull-stripped
IDEAS	1,035	T1, FLAIR; defaced
ARC	2,029	T1, T2, DWI; defaced
MBSR	1,023	T1, DWI; defaced
UCLA LA5c	789	T1, DWI; defaced
QTAB	6,760	minIP, GRE, SWI; defaced
NKI	4,502	T1, T2, DWI; defaced
AOMIC ID1000	4,099	T1, DWI; defaced

The dataset’s heterogeneity is explicit along several axes. First, it combines clinical and research-grade images. Second, it spans multiple scanners, vendors, field strengths, and acquisition protocols. Third, it includes substantial anatomical and pathological diversity, including gliomas, metastases, meningiomas, stroke, aphasia, epilepsy-related surgical cases, aging-related cohorts, and scans with large anomalies. The challenge paper reiterates that FOMO60K is “highly heterogeneous in terms of subject demographics, scanner vendors and models, field strengths, acquisition protocols, and MRI sequence types” (Munk et al., 13 Apr 2026).

This composition has methodological consequences. It makes FOMO60K substantially closer to heterogeneous clinical archives than narrowly curated research datasets, but it also means that nuisance variation is preserved rather than normalized away. A plausible implication is that representation learning methods pretrained on FOMO60K are pressured to learn invariances that are directly relevant to cross-site and out-of-domain deployment.

3. Data format, organization, and preprocessing

All released scans are provided in NIfTI-compressed format (.nii.gz). Subjects are organized as sub_X, sessions as ses_Y, and files within sessions are named by sequence type where metadata are available, such as t1.nii.gz or flair.nii.gz; repeated acquisitions are enumerated, and scans with unavailable sequence metadata may appear as scan_X.nii.gz. Subject identifiers were shuffled to prevent straightforward remapping to original cohort IDs (Munk et al., 17 Jun 2025).

The preprocessing pipeline is intentionally minimal and harmonized rather than aggressively standardized. All scans were reoriented to RAS orientation using mri_coreg from FreeSurfer 7.4.1 with default parameters. Within each session, scans were affinely co-registered to the highest spatial resolution scan, which aligns multi-sequence data without warping images into a common atlas space. For diffusion MRI, 4D data were converted into structurally usable 3D volumes by extracting b=0 when available and by selecting and averaging b=1000 directions most closely aligned with the canonical axes using cosine similarity to $\mathbf{e}_x$, $\mathbf{e}_y$, and $\mathbf{e}_z$ (Munk et al., 17 Jun 2025).

Face removal and de-identification combine pre-existing defacing or skull-stripping from source datasets with additional skull-stripping using SynthSeg when needed. Skull-stripping was applied to images not already skull-stripped or defaced, or when visual inspection detected residual cranial features that could compromise anonymization. The descriptor also states what was not done: there is no normalization to a common template such as MNI, no global intensity normalization or histogram matching, no dataset-level cropping or padding to fixed dimensions, and no advanced bias-field correction pipeline is described at the FOMO60K level (Munk et al., 17 Jun 2025).

These choices preserve raw scanner- and protocol-specific characteristics. That design is deliberate: FOMO60K is not a homogenized atlas-space resource, but rather a standardized release of heterogeneous volumetric data suitable for SSL pipelines that can tolerate variable voxel spacings, matrix sizes, and modality availability.

4. Role in self-supervised learning and the FOMO25 challenge

FOMO60K’s central research role is as a pretraining corpus for SSL-based brain MRI foundation models. In the FOMO25 challenge, the method track constrained participants to pretrain only on FOMO60K, whereas the open track allowed any pretraining data. Downstream evaluation was performed on three clinical tasks: infarct classification, meningioma segmentation, and brain-age regression, under few-shot and out-of-domain conditions (Munk et al., 13 Apr 2026).

The challenge paper organizes pretraining strategies into local objectives such as masked autoencoding, global objectives such as contrastive learning, and hybrid objectives combining reconstruction and contrastive terms. Representative method-track submissions included MAE-based models, SimMIM-style masked image modeling, MoCo v2-style multimodal contrastive learning, and hybrid VAE-plus-contrastive systems. The paper also reports diverse augmentation and sampling choices on FOMO60K, including random flips, rotations, scaling, elastic deformations, patch sampling between $64^3$ and $160^3$, Gaussian or Rician noise, blur, sharpening, gamma transforms, bias field, and Gibbs-ringing or resampling-artifact simulation (Munk et al., 13 Apr 2026).

The main empirical findings are dataset-level rather than model-specific. Sixteen of nineteen evaluated foundation models outperformed the Supervised-OOD baseline trained from scratch on the same few-shot data, and two method-track models pretrained only on FOMO60K outperformed even the Supervised-ID baseline trained on larger in-domain Danish data. The challenge also reports that no single pretraining objective benefits all tasks: local MAE-like objectives tended to favor segmentation, while hybrid reconstruction-contrastive objectives favored classification and regression. In addition, improvements from scaling model size and training duration did not yield reliable benefits (Munk et al., 13 Apr 2026).

These findings are important because they detach the value of FOMO60K from a single winning architecture. The evidence suggests that FOMO60K is large and heterogeneous enough to support substantial gains under clinical domain shift, but that objective design and adaptation strategy remain decisive.

5. Use of the NKI cohort for MRI super-resolution

Although FOMO60K was created primarily for SSL, later work used its NKI cohort as the sole MRI source for training and evaluating elucidated diffusion models for brain MRI super-resolution. That study describes the NKI cohort as providing T1-weighted structural brain MRI volumes from over 1,300 subjects, each acquired at approximately 1 mm isotropic resolution and stored in NIfTI format. The authors trained on 59 subjects (100 scanning sessions) and tested on 5 subjects (6 sessions, 993 sagittal slices), with a subject-level split to prevent leakage (Chiche et al., 15 Mar 2026).

The super-resolution formulation used synthetic low-resolution data derived from FOMO60K high-resolution volumes. Volumes were intensity-normalized to $[0,255]$ using the 1st--99th percentile range, sliced along the sagittal axis, and each 2D slice was downsampled by a factor of 2 using block averaging to produce a $128\times128$ low-resolution input paired with a $256\times256$ high-resolution target. For 3D training, low- and high-resolution volume pairs were stored directly and random patches of size $32^3$ in LR and $32\times64\times64$ in HR were extracted during training. All tensors were then normalized to $\mathbf{e}_y$0 (Chiche et al., 15 Mar 2026).

Within this setting, the paper compared a full 3D convolutional U-Net with volumetric patches and multi-head self-attention against a 2.5D slice-conditioned U-Net that super-resolves each slice independently while conditioning on one adjacent slice. On the held-out test set, the 3D EDM achieved 37.75 dB PSNR, 0.997 SSIM, and 0.020 LPIPS, outperforming both an off-the-shelf pretrained EDSR baseline (35.57 dB, 0.024 LPIPS) and the 2.5D EDM (35.82 dB PSNR) under the same test data and degradation pipeline (Chiche et al., 15 Mar 2026).

The study is also explicit about scope limitations. Its evaluation is confined to a single cohort within FOMO60K, uses image-domain block-average downsampling, and does not model k-space truncation, noise, or motion artifacts typical of clinical acquisition. The paper therefore treats the NKI subset as a clean, standardized high-resolution source rather than as a full proxy for heterogeneous clinical low-resolution MRI (Chiche et al., 15 Mar 2026).

6. Limitations, ethics, access, and future directions

FOMO60K has several documented limitations. Because it aggregates only public datasets, its demographic and disease distributions are not guaranteed to match real clinical prevalence. The dataset paper notes likely underrepresentation of some geographic, racial, and ethnic groups, uneven age distributions across cohorts, and technical imbalance across modalities, with some sequences much more frequent than others. It also highlights metadata gaps, especially for MGH Wild, where many scans have unknown sequence labels (Munk et al., 17 Jun 2025).

Ethically, the dataset is constructed from public de-identified data. De-identification relies on a combination of existing defacing or skull-stripping, additional SynthSeg-based skull-stripping where needed, and manual inspection for residual cranial features. The MRI super-resolution study likewise states that it uses de-identified human brain MRI data from the publicly available FOMO60K/NKI dataset and that the work involves only secondary analysis of anonymized data (Munk et al., 17 Jun 2025, Chiche et al., 15 Mar 2026).

Access and licensing are described across the companion publications. The dataset descriptor lists the Hugging Face release at https://huggingface.co/datasets/FOMO25/FOMO-MRI, while the FOMO25 challenge paper lists https://huggingface.co/datasets/FOMO-MRI/FOMO60K. The challenge paper states that the FOMO60K pretraining dataset is released under CC BY-NC-SA 4.0, with constituent datasets retaining their original licenses; users are therefore expected to respect both the aggregate release conditions and the license terms of the source cohorts (Munk et al., 17 Jun 2025, Munk et al., 13 Apr 2026).

Future work around FOMO60K is already sketched in the literature. The FOMO25 paper frames it as first-edition infrastructure for future challenge rounds and further work on 3D self-distillation, transformer-based approaches, and better balancing of local and global SSL objectives. The MRI super-resolution paper identifies scaling to the full FOMO60K dataset ($\mathbf{e}_y$1 subjects) as a concrete future direction (Munk et al., 13 Apr 2026, Chiche et al., 15 Mar 2026).

Taken together, these publications establish FOMO60K as more than a static data release. It is a heterogeneous pretraining substrate, a benchmark infrastructure for foundation models under domain shift, and a reusable source of volumetric MRI for downstream tasks that require large-scale, minimally preprocessed brain imaging data.