henhao
Sign in
BioinformaticsNature Methods
Peer-reviewed journal article

FMRIPrep: a robust preprocessing pipeline for functional MRI

Oscar EstébaniD, Christopher Johnson MarkiewicziD, Ross W. BlairiD, Craig A. MoodieiD, Ayse Ilkay IsikiD, Asier ErramuzpeiD, James Doliver KentiD, Mathias GoncalvesiD, Elizabeth DuPréiD, Madeleine Snyder, Hiroyuki OyaiD, Satrajit GhoshiD, Jessey WrightiD, Joke DurneziD, Russell Alan PoldrackiD, Krzysztof J. GorgolewskiiD

Abstract

Preprocessing of functional MRI (fMRI) involves numerous steps to clean and standardize data before statistical analysis. Generally, researchers create ad hoc preprocessing workflows for each new dataset, building upon a large inventory of tools available for each step. The complexity of these workflows has snowballed with rapid advances in MR data acquisition and image processing techniques. We introduce fMRIPrep, an analysis-agnostic tool that addresses the challenge of robust and reproducible preprocessing for task-based and resting fMRI data. FMRIPrep automatically adapts a best-in-breed workflow to the idiosyncrasies of virtually any dataset, ensuring high-quality preprocessing with no manual intervention. By introducing visual assessment checkpoints into an iterative integration framework for software-testing, we show that fMRIPrep robustly produces high-quality results on a diverse fMRI data collection comprising participants from 54 different studies in the OpenfMRI repository. We review the distinctive features of fMRIPrep in a qualitative comparison to other preprocessing workflows. We demonstrate that fMRIPrep achieves higher spatial accuracy as it introduces less uncontrolled spatial smoothness than commonly used preprocessing tools. FMRIPrep has the potential to transform fMRI research by equipping neuroscientists with a high-quality, robust, easy-to-use and transparent preprocessing workflow which can help ensure the validity of inference and the interpretability of their results.

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
3,908
FWCI
110.16
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (24 Jul 2018) · CC BY 4.0

fMRIPrep, a BIDS-driven, self-adapting fMRI preprocessing pipeline, earned above-acceptable expert quality ratings on 50 of 54 OpenfMRI datasets and introduced less smoothing than FSL feat.

fMRI preprocessing is usually rebuilt ad hoc for each study (one survey found 223 unique workflows across 241 studies), and consortium pipelines such as HCP or UK Biobank only work with their own acquisition protocols. fMRIPrep offers an analysis-agnostic workflow that configures itself automatically and produces transparent visual reports, aiming for consistent quality on arbitrary datasets.

Key results

  1. ShownIn a large, heterogeneous OpenfMRI sample, experts rated nearly all datasets above acceptable preprocessing quality, with no manual intervention.

    Three raters scored 213 visual reports at six quality control points, plus an overall score, on a scale from 0 (unusable) to 3 (excellent). After Phase II, 50 of 54 datasets averaged above 'acceptable'. The remaining 4 were all above 'poor' and at or near 'acceptable'. · Validation Phase II – Quality assurance and…

  2. ShownAdding expert visual checks to continuous integration testing found and fixed failures, and output quality rose measurably between releases.

    Phase I used 30 OpenfMRI datasets with participants hand-picked for low MRIQC quality; 21 issues across 7 datasets were fixed before v1.0.0. Phase II found 12 more issues across 11 datasets, fixed in v1.0.3. With v1.0.0, two datasets rated below 'poor' and four below 'acceptable'. With v1.0.7, none rated below 'poor' and two below 'acceptable'. · Validation Phase I – Fault-discovery testing; Validation Phase II – Quality assurance and…; Figure legends

  3. ShownCompared with a standard FSL feat workflow, fMRIPrep added less uncontrolled spatial smoothness and aligned better with anatomy, especially in regions distorted by magnetic susceptibility.

    On all 'stopsignal' scans of DS000030 (N=257), AFNI 3dFWHMx showed that feat outputs were smoother. fMRIPrep's estimate before smoothing was below 4.0mm despite resampling to 2.0×2.0×2.0 mm, close to the native 3.0×3.0×4.0 mm voxels. Between-subject variability maps in MNI space showed tighter anatomical alignment, notably in orbitofrontal regions. · FMRIPrep improves spatial precision through…; Figure legends

  4. SuggestedThe choice of pipeline had little effect on downstream statistics. First-level activations were very similar across pipelines, and group-level maps differed mainly in sharpness.

    Identical feat first-level models gave similar thresholded count maps for the go vs successful stop contrast. FSL flame OLS group models on two disjoint random subsets, run with and without 5.0mm FWHM smoothing, performed equivalently on agreement metrics from uncorrected p maps and from maps binarized at FDR 5%. Map overlap and Pearson correlation tracked input smoothness. · FMRIPrep improves spatial precision through…; DISCUSSION; Figure legends

  5. ClaimedCompared with afni_proc.py, feat, C-PAC and HCP Pipelines, fMRIPrep stands out for fieldmap-less susceptibility distortion correction of BOLD data, output in both volume and surface spaces, and versioned containers.

    This comparison is qualitative only. The authors state that no other pipeline applies fieldmap-less correction (nonlinear registration to an inverted-intensity T1w image, restricted to the phase-encoding axis) to BOLD images. They also state that C-PAC and feat are volume-only, and that among the alternatives only C-PAC officially supports containers. · Analysis-agnostic: fMRIPrep supports a wide…; Susceptibility distortion correction (SDC) in…; Ensuring reproducibility with hard versioning…

Methods, data and reproducibility

Nipype orchestrates the workflow. It reads Brain Imaging Data Structure (BIDS) inputs and builds anatomical and BOLD sub-workflows from ANTs, FreeSurfer, FSL and AFNI at runtime. Transforms are combined so each volume is resampled once with Lanczos interpolation, and the pipeline outputs confound series and HTML reports. Validation used two phases of expert visual QC inside continuous integration.

  • OpenfMRI Phase I fault-discovery sample; participants hand-picked for low quality by visual MRIQC assessment · human · n = 30 datasets
  • OpenfMRI Phase II quality-assurance sample; randomly selected participants rated through visual reports · human · n = 54 datasets; 213 participants
  • OpenfMRI 'stopsignal' task scans used to compare fMRIPrep with FSL feat · human · n = N=257 participants · DS000030
  • Simultaneous MRI/electrocorticography (ECoG) dataset with BOLD signal dropout near implanted electrodes (Online Methods, Figure S11) · human
Methods & tools
Nipype, ANTs (N4BiasFieldCorrection, antsBrainExtraction.sh, antsRegistration, antsApplyTransforms), FreeSurfer (recon-all, bbregister, mri_robust_template, mri_vol2surf), FSL (mcflirt, flirt, fast, feat, flame), AFNI (3dTshift, 3dFWHMx), Nilearn, ICA-AROMA, CompCor (aCompCor, tCompCor), MRIQC, ICBM 152 Nonlinear Asymmetrical template 2009c; OASIS and NKI skull-stripping templates, Docker and Singularity containers (BIDS-Apps model), Brain Imaging Data Structure (BIDS)
Code
https://github.com/poldracklab/fmriprep
Software & environment
Each release ships Docker and Singularity images that fix the versions of fMRIPrep and all its dependencies (BIDS-Apps model). Validation Phase I ended with release 1.0 and Phase II with release 1.0.8. Spatial normalization targets ICBM 152 Nonlinear Asymmetrical 2009c.
Reproducibility gaps
The text gives no URL for the visual reports, rater scores or OpenfMRI participant lists. The full group-level metrics for the feat comparison are deferred to Online Methods, which are not included in the text.
Limitations (5)
  • Preprocessing quality was judged visually (the reports) and semi-visually (variability and smoothness figures), not with quantitative frameworks. The authors note that existing frameworks such as NPAIRS impose restrictive assumptions on the data and workflow. · stated by the authors
  • Datasets with a reduced field of view were left out of the evaluation because they are not yet fully supported. Very narrow fields of view may lack enough information for standard registration. · stated by the authors
  • The pipeline may underperform for special populations such as infants, or for brains with tumors, resections or lesions. Use with other species, such as rodents or nonhuman primates, would need suitable atlases. · stated by the authors
  • Quantitative benchmarking compares fMRIPrep with only one alternative (FSL feat) on one dataset (DS000030). Advantages over AFNI, SPM, C-PAC or HCP Pipelines are argued only qualitatively. · noted in this summary
  • Three of the authors (CJM, KJG, OE) gave the quality ratings, with no stated blinding. Dataset counts also differ between the text (54 datasets) and the Figure 3 legend (213 participants from 58 studies). · noted in this summary

Who should read:Read this if you preprocess task or resting-state fMRI from varied datasets and want fMRIPrep's design choices, how it was validated, and how it compares with FSL feat.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISEstéban '20Notter '22Waller '22Wang '24Goncalves '25Kristanto '24
83% semantic matchBioinformatics
Oscar Estéban, Rastko Ćirić, Karolina Finc+24 more·Nature Protocols·Jun 8, 2020· 315 citations