Self-supervised pretraining has transformed language and vision, but its application to functional MRI (fMRI) is constrained by data access. The largest standardized fMRI resources are governed by data use agreements and dominated by a small number of acquisition protocols, while the thousands of openly licensed datasets on OpenNeuro are individually small, heterogeneously organized, and unprocessed. We present the first release of Open-fMIND, a uniformly preprocessed, openly accessible corpus of parcellated fMRI timeseries from 322 OpenNeuro studies, comprising 11,379 subjects and 45,731 functional runs. Each run is preprocessed with fMRIPrep and parcellated under 52 atlases spanning 14 families, yielding a massive 2.4 million timeseries. Timeseries are released either denoised with our default pipeline (preproc) or as the fMRIPrep output with its confound files (quasiraw), with functional connectivity matrices, per-run provenance and motion metadata. Open-fMIND is intended as a pretraining resource for self-supervised brain representation learning. Pretraining on it yields downstream performance comparable to that obtained with UK Biobank.
Figures & tables
Figure 1: Open-fMIND pools many small datasets into a corpus larger than any single one of them. (a) The 1,066 OpenNeuro MRI datasets, by size and number of subjects. Blue points are datasets with released runs in Open-fMIND, grey points the others. Dashed lines mark constant size per subject, and the black line is the Pareto frontier (no dataset has both more subjects and more data). The star is the release itself: its 11,379 subjects and the size of its files (4,321 GB, both versions and all atlases). (b) Scan time against subjects for the released datasets, restricted to released subjects and runs.
Table 2: Atlas inventory, grouped by parcellation family. “Hard” atlases assign each voxel to exactly one parcel; “soft” atlases are probabilistic and spatially overlapping. ROI counts are those of the released sidecars ( ROILabels ).
Datasets
Subjects
Runs
OpenNeuro (all fMRI datasets)
1,066
45,067
Retrieved for this release
359
14,016
52,505
Released (excluding failures)
Open-fMIND ( quasiraw )
322
11,379
45,731
Open-fMIND ( preproc )
320
11,320
45,019
Table 3: Datasets, subjects and runs on OpenNeuro, in the data retrieved for this release, and in Open-fMIND.
Figure 2: Runs form regions in connectivity space that differ in motion, dataset of origin and task. (a) UMAP (15 neighbors, minimum distance 0.1) of the Schaefer-400 connectivity vectors of 44,241 preprocessed runs from 288 datasets; datasets with fewer than 5 subjects are left out (32 datasets, 63 subjects, 778 runs). The 12 dense clusters found by HDBSCAN (minimum cluster size 500, minimum samples 50, leaf selection; 36% of runs) are colored and numbered, and the other runs are grey. Boxes mark the clusters enlarged in (b-d), each of which maximizes one criterion. (b) Motion: the cluster with the highest median mean FD (cluster 11, 231 datasets; 0.165 mm against 0.124 mm over all runs), colored by mean FD capped at the 95th percentile. (c) Dataset coverage: the cluster with the highest median dataset coverage (cluster 2, 16 datasets, 31%), colored by dataset. (d) Task-only cluster: the cluster with the lowest share of rest runs (cluster 5, no rest runs, 23 datasets, 34 task names), colored by task name. In (c) and (d), runs of other clusters inside the box are grey, and legends name the six most frequent groups. (e) The 64% of runs outside any dense cluster. (f) Rest (19%) and task (81%) runs. (g) All runs colored by dataset. Each dataset has its own color, set by the median position of its runs in the plot (hue by x, lightness by y), so datasets with nearby centers have similar colors. Axes are arbitrary.
Figure 3: Composition, population coverage and validation experiments. (a,b) Subjects per released dataset (dashed: median) and sessions per subject with released runs. (c,d) Reported sex and age (stacked by sex) of the released subjects. (e-g) Within- and between-network correlation in resting-state runs (Schaefer-400, Yeo 7 networks): mean per network pair (outlined diagonal: within-network), per run (dashed: equality), and per network (shaded: pooled; *** p<0.001 , Holm-corrected). (h,i) Agreement of the finer atlas collapsed onto the coarser one with the native coarser matrix, per run (dashed: medians) and per atlas pair (median and interquartile range). (j-l) Identification across visits: fraction identified within the top k (dashed: chance), similarity to the true and to the best other subject (dashed: means), and accuracy per dataset against chance (dashed: equality; marker area proportional to subjects). (m,n) QC-FC of 79,800 edges by FDR significance (dashed red: median) and against inter-ROI distance (dashed red: least-squares fit). (o-q) Classification of the dataset of origin: balanced accuracy against chance and a permutation null, variance fraction per edge explained by dataset (dashed: median ω2 ), and recall per dataset against its size (dashed: chance). Panels e-q use the preproc version; sample sizes are in the text.
Figure 4: A model pretrained on Open-fMIND reaches accuracy close to that of a model pretrained on UK Biobank. Frozen Brain-Semantoks encoders with a linear or kernel readout on HCP-YA, ABIDE and ADHD-200. Bars are means and error bars standard deviations over cross-validation folds (200 for HCP cognition) or repetitions (20 for HCP sex, 10 for ABIDE and ADHD-200). Accuracies are balanced accuracies, age is predicted as one of 5 classes of equal size, and dotted lines mark chance (50% for two classes, 20% for five).
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Accession
Name
License
Subjects
Runs
ds000001
Balloon Analog Risk-taking Task [ 22 ]
CC0-1.0
16
48
ds000002
Classification learning [ 23 ]
PDDL-1.0
17
102
ds000003
Rhyme judgment [ 24 ]
CC0-1.0
13
13
ds000005
Mixed-gambles task [ 25 ]
PDDL-1.0
16
48
ds000006
ds000006 [ 26 ]
PDDL-1.0
14
162
ds000007
Stop-signal task with spoken & manual responses [ 27 ]
PDDL-1.0
20
118
Appendix
Table 4: OpenNeuro datasets in Open-fMIND v2.0 (322 datasets), with the title, license and citation of the OpenNeuro snapshot that was downloaded (the latest one for the 17 datasets with no recorded snapshot; retrieved 2026-11-10). Citations give the authors, title, snapshot version and, when registered, the DOI. Subjects and Runs are the released subjects and runs of each dataset ( quasiraw , which contains all released runs). Each accession links to its page at https://openneuro.org/datasets/<accession> . Licenses: CC0-1.0 (251), PDDL-1.0 (34), not stated (30), CC-BY-SA-4.0 (3), CC-BY-4.0 (2), CC-BY-NC-4.0 (1), other (1).
Functional magnetic resonance imaging (fMRI) foundation models learn reusable brain representations for prediction, transfer, and neuroscience discovery.
Functional magnetic resonance imaging (fMRI) is a powerful tool for investigating human brain function. However, the high cost of data acquisition and the inherent subjectivity of psychiatric rating scales often lead to datasets with small sample sizes and variable label quality, especially when targeting a specific neurological condition. Combined with the inherently high dimensionality of fMRI data, these limitations substantially increase the risk of model overfitting. Recent years have seen growing interest in developing fMRI foundation models by combining multiple datasets; however, the computational resources needed for pretraining and fine-tuning are often prohibitive. We show that a lightweight self-supervised framework yields representations that generalize across diverse downstream tasks, outperforming fully supervised baselines and approaching the performance of large-scale models. We introduce BrainSimSiam, a data-efficient self-supervised representation learning framework that leverages positive-only data pairs to learn robust and generalizable features. We demonstrate that the learned representations achieve strong performance across multiple downstream classification and regression tasks, highlighting the potential of BrainSimSiam for data-limited neuroimaging applications.
Jiyao Wang, Peiyu Duan, Nicha C. Dvornek +4
Department of Biomedical Engineering, Yale University, New Haven, CT, USA · Radiology & Biomedical Imaging, Yale School of Medicine, New Haven, CT, USA · Electrical Engineering, Yale University, New Haven, CT, USA +1
We study the problem of training self-supervised foundation models for functional MRI. Our main contributions are: (1) we introduce a new model family (CortexMAE) trained using the masked autoencoder framework on 2.1K hours of open fMRI data, and (2) we release the first open evaluation suite (Brainmarks) for fMRI foundation models. Our core innovation is simple: we adapt the Vision Transformer to fMRI by first converting each 3D fMRI volume to a 2D map using a cortical flat map projection. We directly compare flat maps to both parcellation and volume-based representations. While each has its advantages, flat maps generally perform best. We perform the first systematic scaling analysis for fMRI and observe strict power law scaling, albeit with limits. Finally, we use Brainmarks to do controlled benchmark comparisons. On subject-level trait prediction, we report a challenging null result: no single model achieves clear state-of-the-art performance. Moreover, all models struggle to outperform a simple functional connectivity baseline. On cognitive state decoding, we observe more robust performance, and in this setting our CortexMAE family outperforms prior models by a large margin. Code, models, and datasets are available at https://github.com/MedARC-AI/CortexMAE and https://github.com/MedARC-AI/Brainmarks.
Functional magnetic resonance imaging (fMRI) commonly characterizes large-scale brain organization through functional connectivity (FC). However, an observed FC matrix is not a fixed brain graph but a statistical estimate determined by the finite BOLD frames used to compute it. Conventional FC-based approaches typically represent each scan with a single Pearson-correlation matrix, leaving frame-resolved variations in regional interactions unexploited. To address this limitation, we introduce CoFlu-Brain, which organizes framewise cofluctuation observations into multiple retained and dropped views, each summarized as an FC matrix. The model learns connectivity structure reproducible across retained views and predicts the representations of the corresponding dropped views, while adopting SIGReg from LeJEPA to prevent representation collapse. We pretrained CoFlu-Brain on 13,444 subjects from eight dataset groups and evaluated its representations via frozen-encoder linear probing on seven classification and two age-prediction tasks across seven datasets. CoFlu-Brain showed strong transfer across demographic, developmental, and clinical prediction tasks, outperforming the evaluated released models on most tasks. Random frame partitioning improved aggregate transfer, while the complete objective achieved the strongest overall performance. Code is available.
Ziteng Sui, Junfeng Xia, Wenhao Ye +2
Department of Biomedical Engineering, Southern University of Science and Technology, China · School of Biomedical Engineering, Shenzhen University, China · Department of Computer Science, University of Warwick, UK