The NiPreps ecosystem for reproducible neuroimaging

Russ Poldrack

Stanford University

The standard design (anti-)pattern for fMRI preprocessing

• Pick a single software package
  • Usually based on considerations other than performance

The standard design (anti-)pattern for fMRI preprocessing

• Pick a single software package
  • Usually based on considerations other than performance
• String together the tools from that package into a script to run the preprocessing workflow

The standard design (anti-)pattern for fMRI preprocessing

• Pick a single software package
  • Usually based on considerations other than performance
• String together the tools from that package into a script to run the preprocessing workflow
  • Generallly written by a student or postdoc with little software engineering experience
• Hope it keeps working over time…
• Trust that it does the right thing…

Towards a new design pattern for preprocessing: fMRIPrep

• A robust workflow for preprocessing fMRI data
  • Tested on a large number of fMRI datasets from OpenNeuro
  • Adapt to each dataset and processes it in the best way possible
  • Outputs to BIDS-Derivatives format
  • Provides powerful visualizations :
    • to help identify problems
    • allowing glass-box access to training researchers
  • Currently averaging ~5000 successful runs per week

Esteban et al., 2019, Nature Methods

RF1MH121867: NiPreps (NeuroImaging PREProcessing toolS)

… the overarching objective of this project is to develop NiPreps, a software framework to perform standardized preprocessing of diverse neuroimaging data.

• Aim 1: solidify the foundations of the NiPreps integration.
• Aim 2: enable integrative analysis approaches of heterogeneous data.
• Aim 3: accelerate the dissemination of NiPreps to the neuroscience community through hackathons and “docusprints”.

Why?

In developing fMRIPrep, we learned about critical aspects of neuroimaging workflows. NiPreps is envisioned as a generalization of fMRIPrep.

RF1MH121867: Sites

Poldracklab (Stanford)

Esteban Lab (CHUV Lausanne)

Satterthwaite Lab (Penn)

Milham Lab (Child Mind Inst)

Rokem Lab (UW)

The NiPreps community

www.nipreps.org

Aim 1: Developing and refining reusable infrastructure/middleware components

• TemplateFlow: FAIR Sharing and management of neuroimaging templates and atlases
• SDCflows: Integrating susceptibility distortion correction (SDC)
• NiReports: A modular visual reports system
• NiTransforms: Spatial transforms integration

TemplateFlow: FAIR sharing of neuroimaging templates and atlases

• Templates and atlases are commonly used in neuroimaging research

• There is significant lack of clarity in the use of these templates

  • There are numerous versions of the widely used “MNI template”

• Templateflow provides programmatic access to a database of templates and mappings between them

• Easy to use for humans and machines:

Ciric et al., 2022, Nature Methods

SDCFlows: Susceptibility Distortion Correction workflows

• SDCflows aims to provide a unified interface to susceptibility distortion correction methods
  • Defines a shared representation model (B-Spline) for the field map
  • “decouples” estimation and application steps (increasing modularity)
• Overhaul started early 2021 (Esteban et al., OHBM 2021)
  • Faced many technical challenges
    • Requiring numerous bugfixes and “edge” cases
  • Developed new educational materials & Jupyter notebooks

NiReports: New visualization tools from MRIQC

(Provins et al., ISMRM 2022)

• MRIQC is a quality control workflow for structural/functional MRI
• Developing a number of visualizations that will go into NiReports
  • Added visualization of voxels at the edge of the brain (“crown”)
  • Added hierarchical sorting of rows (voxels) to enhance patterns (Aquino et al. 2019)

Infrastructure: Architectural redesign

• Problem: fMRIPrep’s “one size fits all” design has limitations for emerging use cases
  • Archiving preprocessing results requires balancing storage costs against possible use cases.
  • Including alternative algorithms requires custom code to integrate.
• Solution: Accept pre-computed derivatives and defer computationally cheap operations
  • E.g., Deep learning segmentations and masks can be accepted, skipping fMRIPrep defaults.
  • Multiple template registrations can be archived, analysts may resample BOLD series to different spaces on demand.
• This approach is implemented in SDCFlows and is being generalized to other components.

Aim 2: Expand the portfolio of end-user NiPreps

• ASLPrep
• dMRIPrep
• PETPrep
• fMRIPrep-infants (aka NiBabies)
• fMRIPrep-rodents (aka NiRodents)

Workflows: ASLPrep (cerebral blood flow quantification)

• A robust workflow for preprocessing arterial spin labeling (ASL) data
  • Including cerebral blood flow (CBF) quantification
  • Provides quality evaluation for CBF maps
  • Provides CBF quantification at the regional level using atlases

Adepimbe et al., 2022, Nature Methods

Workflows: dMRIPrep (diffusion MRI)

• A workflow for preprocessing of diffusion MRI data
• Development currently focused on eddymotion
  • an algorithm to estimate head-motion (modality-agnostic) and modality-specific artifacts (eddy currents in the case of dMRI)

Preprint: Pisner et al., 2022.

Workflows: PETPrep (positron emission tomography)

• A NiPreps workflow for PET preprocessing
  • Successfully merged petsurfer into nipype (1.8.0)
  • Incorporated nipype implementation of a robust head motion correction workflow (petprep_hmc)
  • Developing a BIDS-Derivatives standard for PET derivatives
  • Model-based head-motion correction leveraging eddymotion in progress

Martin Norgaard, in prep

Workflows: fMRIPrep for infants

• Collaboraration with Damien Fair & HBCD team
• New developments
  • Support for pre-computed derivatives (mask, segmentations).
  • Improved robustness and validity of CIFTI-2 outputs.
• Upcoming developments
  • Morphometric outputs (cortical thickness, curvature)
  • Improvements to susceptibility distortion correction versatility
  • T2 assisted surface generation

Mathias Goncalves, in prep

Aim 3: Consolidate the NiPreps community

• Project monitoring infrastructure: MIGAS
• Evaluation of cross-workflow reproducibility
• Hackathons and documentation
• Best practices and educational resources
• NMIND: Building common standards for software development

Project monitoring: MIGAS

• An open-source, customizable telemetry solution
• Allows collecting usage information, errors, and status throughout a process’s lifetime
• Easy to deploy with various cloud providers (Heroku / GCP / AWS)
• Available as a Python package: https://pypi.org/project/migas/

Mathias Goncalves, in prep

Reproducibility: Cross-workflow evaluation

• CMI team developed a CPAC implementation of fMRIPrep
  • Able to achieve high levels of reproducibility in connectivity metrics between harmonized workflows
  • Helped identify causes of divergence, such as use of different versions of MNI template

Xinhui Li et al., under review

Hackathons and documentation

• Held a hackathon/documentation sprint in Glasgow following OHBM 2022
• Will participate in Brainhack Global 2022
• Planning to hold a hackathon/documentation sprint in Montreal in association with OHBM 2023

Best practices and educational resources

• Collaborative QC-Book (educational, ISMRM 2021): https://nipreps.org/qc-book

• MRIQC-SOPs (standard operating procedures)

  • A GitHub template-repository to create and maintain versioned SOPs documentation and checklists.
  • Example: https://nipreps.org/mriqc-sops/

• MRIQC Protocol report (Hagen et al., in preparation)

• Frontiers’ research topic on QC of fMRI (Provins et al., under review)

• Biases introduced by defacing in QC (Provins et al., pre-registered report under review)

NMIND: Building common standards for software development

• NMIND: Nevermind, this Method Is Not Duplicated
  • Alignment: development and adoption of standards for critical software component
  • Testing: accessible and (semi-)automated mechanisms for evaluating standards compliance
  • Engagement: widespread promotion and adoption of the NMIND collaborative standards

Thank you!

fMRIPrep usage

• Usage tracked using an opt-out telemetry system
  • Allows quick identification of bugs and usage patterns
• Currently averaging ~5000 successful runs per week

Note

We are developing an open-source alternative called “migas”, to replace Sentry

SDCFlows in a nutshell

• What? Wraps or implements methods for estimating field maps in various scenarios:
  • covering (i) so-called TOPUP, (ii) phase-difference fieldmaps and their variants, (iii) fMRIPrep’s “fieldmapless”; but
  • it does not cover point-spread-function approaches (extremely marginal).
• How?
  • Defines a shared representation model (B-Spline) for the field map (comparability ↑↑, methodological variability ↓↓)
  • Technical perk: “decouples” estimation and application steps (modularity ↑↑, methodological variability ↑).
• Why?
  • A single tool can be applied to correct for distortion, no matter how it was estimated (method comparability ↑↑)
  • The model coefficients can easily be moved with head motion (e.g., dynamic fieldmap)
  • The model coefficients can easily be integrated in spatial transformation chains (one-shot interpolation)
  • The model coefficients can easily be integrated within other software (e.g., eddymotion)

Result YR1: NiRodents

• MRIQC-rodents saw a first release presented in EMIM 2022
• fMRIPrep-rodents is currently integrating new SDCFlows’ API
• NiRodents has stimulated several improvements of the reporting system (see next)
• NiRodents has stimulated the inclusion (and revision of existing) rodent templates in TemplateFlow

Result YR1: new visualizations in MRIQC and fMRIPrep

(Provins et al., ISMRM 2022)

• Added visualization of voxels at the edge of the brain (“crown”)
• Added hierarchical sorting of rows (voxels) to enhance patterns (Aquino et al. 2019)
• fMRIPrep: added “edge” regressors (Patriat et al. 2008)

Results YR1: Other areas

• NiReports: provides standard mechanisms to build “reportlets” and full reports.
  • A repository has been initiated
  • Code already exists, but it is scattered across tools (e.g., MRIQC, fMRIPrep)
  • In the OHBM Hackathon, work initiated for moving MRIQC visual components into it (fMRIPrep will follow).
• NiTransforms:
  • The component is in a stable status
  • The component is key to achieving a redesign of fMRIPrep (described next)