Pipeline Overview

Diffusion preprocessing follows a common sequence across large studies: brain extraction, denoising, Gibbs ringing removal, susceptibility distortion correction, head motion and eddy current correction, and model fitting. The Human Connectome Project, UK Biobank, and ABCD pipelines all implement that sequence. The configuration documented here is one version of it.

Maximov, Alnaes, and Westlye (2019) quantify which of these corrections change diffusion metrics and by how much: Human Brain Mapping 40(14), 4146-4162, doi:10.1016/j.neuroimage.2021.118756.

Pipeline Stages

Twelve steps in two parts. Part A applies the corrections required of any diffusion dataset. Part B fits the diffusion models and writes the files a tractography workflow reads. Several further steps are useful but sit outside the required path and are listed separately below.

The number of steps a given study needs depends on its acquisition and its analysis. This is one tested configuration rather than the only valid one.

Part A — Core Preprocessing

Part A has two routes. Both take raw data to corrected, brain-masked, T1-aligned diffusion data, and both feed Part B.

Route 1: QSIPrep

Step	Name	Purpose	Tool(s)
—	QSIPrep	Skull stripping, TOPUP, denoising, Gibbs removal, eddy, bias correction, and T1 alignment in one containerized run, with a QC report per participant	QSIPrep (Docker / Singularity)

QSIPrep reads BIDS, so DICOM conversion and BIDS layout happen first; the route page covers that. Because it writes the DWI into the T1 grid, Step 10 is not needed afterwards.

Route 2: Manual Steps

Step	Name	Purpose	Tool(s)
1	DICOM to NIfTI	Standardize raw scanner data	dcm2niix
2	Skull Stripping	Remove non-brain tissue	ANTs
3	B0 Concatenation	Merge AP/PA fieldmaps	FSL
4	TOPUP	Correct susceptibility distortions	FSL
5	Mean B0	Create high-SNR reference image	FSL
6	Brain Masking	Define analysis region	FSL
7	Denoising & Gibbs	Remove noise and ringing artifacts	MRtrix3
8	Eddy Correction	Correct motion and eddy currents	FSL

Choosing a Route

	Route 1: QSIPrep	Route 2: Manual
Suits	New studies, multi-site studies, anyone who wants a pinned and reproducible container	Learning what each correction does; acquisitions QSIPrep does not support; full control of every parameter
Input	BIDS tree (conversion done first)	Raw DICOM
Runs as	dcm2niix into BIDS, then one command per participant	Eight scripts per participant
Tools inside	FSL topup and eddy, MRtrix3 dwidenoise and mrdegibbs; SynthStrip for skull stripping; N4 bias correction on the DWI by default	ANTs for skull stripping; no DWI bias correction
T1 ↔ DWI	DWI resampled into the T1 grid; Step 10 skipped	Separate grids; Step 10 computes the transform
QC	One HTML report per participant	A check section on each step's page
Output	BIDS derivatives, linked into the required layout by the route page's script	The required layout directly

The two routes are not verified to be numerically identical. Most steps use the same tool, but skull stripping differs, QSIPrep's Gibbs removal is off unless requested, and its bias correction is on unless disabled; the route page lists these step by step. Do not mix routes within one study.

Part B — Tractography Readiness

Steps 11 and 12 take two forms. Multi-shell data uses three-tissue MSMT-CSD. Single-shell data uses single-tissue CSD, given in Step 12. Both produce the same output file.

Step	Name	Purpose	Tool(s)
9	Tensor Fitting	Compute FA, MD, RD, AD	FSL
10	Registration	Align T1 to diffusion space	FSL
11	Response Functions	Estimate per-tissue reference signals	MRtrix3
12	FOD Estimation	Deconvolve to fiber orientation distributions	MRtrix3
—	Outputs for Tractography	Verify outputs before tracking	—

Optional Steps

These are covered in this tutorial but are not required to reach tractography.

Name	Purpose	When it applies
BedpostX	Per-voxel fiber orientations for FSL probtrackx2 — the FSL counterpart to Steps 11–12, which do the same job for MRtrix3 tckgen	Only if your tractography tool is probtrackx2; pick one, not both
Shell Extraction	Isolate b-value shells	Tensor fitting — not for MSMT-CSD, which needs all shells
ICV Calculation	Estimate intracranial volume	As a statistical covariate
BIDS & pyAFQ	Organize for pyAFQ	pyAFQ's automated whole-brain bundle recognition

---

Pipeline Flowchart

Steps 9 and 11 both branch from Step 8, or from QSIPrep output, which replaces Steps 1–8 entirely. Tensor fitting and FOD estimation are independent of each other and can run in parallel. Registration (Step 10) needs the skull-stripped T1 from Step 2 and is skipped on the QSIPrep route. Everything converges at the required outputs.

---

Rationale

Core Operations (Required)

The following stages are accepted across the field as minimum required operations for diffusion MRI preprocessing:

• Brain extraction (Step 2) -- non-brain tissue introduces artifacts and errors in all downstream steps
• Susceptibility distortion correction (Step 4) -- corrects geometric distortions caused by magnetic field inhomogeneities
• Denoising (Step 7) -- thermal noise degrades tensor estimation, especially at higher b-values
• Motion and eddy current correction (Step 8) -- subject motion and eddy currents cause volume misalignment and signal distortions
• Tensor fitting (Step 9) -- the fundamental computation that produces FA, MD, and other diffusion metrics

Software-Specific Steps

Some stages exist because of specific software requirements rather than conceptual necessity:

• B0 concatenation (Step 3) -- FSL's topup requires AP and PA B0 volumes in a single file
• Mean B0 creation (Step 5) -- averaging multiple B0 volumes produces a higher-SNR reference for brain masking
• Brain masking (Step 6) -- eddy requires an explicit brain mask; this refines the initial skull strip

Processing Standards

This sequence reflects the processing standards established by:

• Human Connectome Project (HCP) -- the gold standard for diffusion MRI processing
• UK Biobank -- large-scale population imaging with automated preprocessing
• ABCD Study -- multi-site developmental neuroimaging with rigorous QC

---

Directory Structure Convention

Organizing outputs into a consistent directory structure makes the pipeline manageable across subjects. The recommended layout:

project/
  raw/                    # Raw DICOMs
  NIFTI/                  # Converted NIfTI files
  derivatives/
    ANTs/                 # Skull stripping outputs
    TOPUP/                # Distortion correction
    EDDY/                 # Motion/eddy correction
    DTIFIT/               # Tensor fitting outputs
    FLIRT/                # Registration matrices
    ICV/                  # Intracranial volume
    pyAFQ/                # BIDS-formatted for tractography
  config/                 # Configuration files (acqp.txt, index.txt)

Each subject's outputs live within the appropriate derivatives subdirectory. For example:

derivatives/
  DTIFIT/
    sub-001/
      sub-001_dtifit_FA.nii.gz
      sub-001_dtifit_MD.nii.gz
      sub-001_dtifit_RD.nii.gz
      sub-001_dtifit_AD.nii.gz
    sub-002/
      ...

This structure keeps raw data separate from processed outputs and groups outputs by processing stage.

---

Reference Implementation

The full pipeline described in this tutorial has been implemented and run end to end on a real multi-shell dataset:

Repository: github.com/DiffusionTensorImaging-Repos/SDN-IMPACT-DTI

It includes batch processing scripts, configuration files, and the QC output from each step. Use it as a reference when adapting the pipeline to your own data — see the Worked Example page for what it contains.

---

Getting Practice Data

The Practice Data page lists sample datasets suitable for working through this tutorial.