Open-fMIND: A Large-Scale Open-Access fMRI Corpus for Pretraining Brain Foundation Models

Carlo Alberto Barbano, Giovanni Marraffini, Demian WassermannarXiv 2026 · arXiv

TL;DR

Open-fMIND is a massive, open-access fMRI corpus designed to fuel the next generation of brain foundation models. By standardizing and preprocessing data from over 11,000 subjects across 322 diverse studies, it provides a "model-ready" alternative to restricted, gated datasets like the UK Biobank.

What Problem Does This Paper Address?

In fields like NLP and computer vision, foundation models thrive on massive, diverse datasets. Functional MRI (fMRI) research, however, faces a "data bottleneck." While large-scale standardized resources exist (e.g., the UK Biobank or HCP), they are often "gated," requiring complex data use agreements that hinder open science and rapid iteration.

Conversely, open platforms like OpenNeuro host thousands of datasets, but they are a "long tail" of small, heterogeneously organized, and unprocessed studies. The effort required to retrieve, clean, and harmonize hundreds of these independent studies has kept this data largely untapped for large-scale machine learning. Open-fMIND bridges this gap by doing the heavy lifting of preprocessing and standardizing this "long tail" into a single, massive, open-access corpus.

Methods

The authors developed a uniform pipeline to transform raw OpenNeuro data into parcellated region-of-interest (ROI) timeseries.

  1. Acquisition: They queried the OpenNeuro API, identifying over 1,000 datasets and selecting 322 for this first release.
  2. Preprocessing: Every functional run was processed using fMRIPrep, a gold-standard tool for motion correction and anatomical alignment.
  3. Parcellation: Instead of picking one "best" brain map, they parcellated each run under 52 different atlases (spanning 14 families). This allows researchers to study how different assumptions about brain organization affect model performance.
  4. Denoising: Data is released in two flavors: quasiraw (fMRIPrep output with confound files) and preproc (denoised using a standard pipeline including regression of 26 confounds and signal scrubbing).
  5. Connectivity: They also provide functional connectivity matrices for every run and atlas.
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.

Datasets

The first release of Open-fMIND includes:

  • 11,379 subjects and 45,731 functional runs.
  • Data sourced from 322 OpenNeuro studies.
  • 429 distinct task labels, offering significantly more functional diversity than resting-state-heavy gated resources.
  • 2.4 million parcellated timeseries in total.

For evaluation, the authors compared models pretrained on Open-fMIND against those pretrained on the UK Biobank and tested them on downstream benchmarks like HCP-YA, ABIDE, and ADHD-200.

Key Insights and Ideas

  • Diversity as a Feature: Unlike the UK Biobank, which is dominated by a few acquisition protocols, Open-fMIND is highly heterogeneous. It captures a vast array of cognitive tasks and scanner types, which may lead to more robust and generalizable brain representations.
  • The "Atlas Axis": By providing 52 different parcellations, the authors treat the choice of atlas as a "view" or "augmentation." This opens the door for contrastive learning techniques where a model learns by comparing different ways of looking at the same brain activity.
  • Clustering by Task and Motion: Using UMAP embeddings of connectivity matrices, the authors show that runs naturally cluster by task type and, unfortunately, by the amount of head motion (Figure 2). This highlights the importance of rigorous denoising in large-scale fMRI aggregation.
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.

Key Results

The headline result is that open data can match gated data for pretraining.

The authors pretrained a foundation model (Brain-Semantoks) on Open-fMIND and found that its performance on downstream tasks—such as predicting sex or cognitive factors—was virtually identical to a model pretrained on the restricted UK Biobank. On the HCP-YA dataset, both models achieved an AUROC of 0.945 for sex prediction and a Pearson correlation of ~0.41 for cognition, significantly outperforming random initialization.

Performance comparison showing that Open-fMIND pretraining matches or nears the performance of UK Biobank pretraining across multiple downstream tasks.

Limitations and Future Work

  • Parcellated Release: To keep the dataset size manageable (4.3 TB), only ROI timeseries are released. Researchers needing voxel-wise detail would need to rerun the pipeline on the raw data.
  • Site Effects: Because the data comes from hundreds of different scanners, "site effects" are present. The authors provide metadata to help users account for this, but harmonization remains a challenge for downstream users.
  • Expansion: This is "Release 1." The authors plan to extend the corpus to the full set of available OpenNeuro data in future iterations.
  • Label Curation: The authors intentionally did not harmonize phenotypic labels (like age or clinical diagnosis) across studies, leaving this complex task to researchers focused on specific supervised downstream applications.

Summary updated

Abstract

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.
BibTeX
@misc{barbano2026open,
  author = {Carlo Alberto Barbano and Giovanni Marraffini and Demian Wassermann},
  title = {{Open-fMIND: A Large-Scale Open-Access fMRI Corpus for Pretraining Brain Foundation Models}},
  year = {2026},
  eprint = {2611.21215},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21215},
}
Figure