Data requirements¶

DELTA-SVD analyses multi-directional diffusion MRI that has already been preprocessed. It does not perform image preprocessing itself. The guidance below describes acquisitions and preprocessing that make data well suited to the pipeline. It reflects current best practice rather than a single mandated protocol: sensible deviations are fine, but these recommendations are a safe default.

Required input¶

Each timepoint is described by four files. The DWI must already be preprocessed (see Preprocessing below); DELTA-SVD does not preprocess it for you.

• DWI image — a 4D, preprocessed diffusion-weighted NIfTI (.nii.gz or .nii).
• b-values — an FSL-format bval file matching the DWI.
• b-vectors — an FSL-format bvec file matching the DWI.
• Brain mask — preferably a binary brain-mask NIfTI (.nii.gz or .nii) in DWI space (aligned with the DWI image). DELTA-SVD binarises masks on input: values greater than zero become 1; zero and negative values become 0.

A longitudinal run takes one such set per timepoint. For how to pass these on the command line (and the naming convention that lets the bval/bvec/mask be inferred from the DWI path), see Usage.

Acquisition¶

• Field strength — 3T is recommended.
• Resolution and directions — a recommended minimum of 2 mm isotropic voxels and at least 20 diffusion-weighting directions. Higher resolution and more directions improve the tensor estimates, at the cost of scan time. DELTA-SVD refuses data that leaves it fewer than 12 distinct directions after the b-value selection, and warns below the recommended 20: the free-water fraction is estimated against the residual of the tensor fit, so it needs angular sampling well beyond the six directions the tensor itself requires. Repeated acquisitions of the same direction average noise but add no further constraint, and are counted once.
• b-value — include at least one shell suitable for diffusion-tensor fitting, around b = 1000 s/mm². By default DELTA-SVD fits the tensor on b-values in the range 800–1200 s/mm² (together with the b ≈ 0 volumes) and automatically selects those shell(s) from the data, so multi-shell acquisitions are fine, and the appropriate shell is picked for you. A different selection can be made with --bRange or --shells, but only within 250–1800 s/mm². That window is a deliberately wide outer bound: it marks where the diffusion tensor model breaks down, and is not a suggestion to move away from b ≈ 1000 (see Advanced usage).
• Direction scheme — sample the diffusion directions over the whole sphere rather than a half-sphere (a custom whole-sphere direction set is recommended). Full-sphere sampling gives eddy-current and motion correction (FSL eddy) better-conditioned data to work with.
• Reversed phase encoding — acquire at least one b ≈ 0 volume (or a short series) with reversed phase-encoding direction, so that susceptibility-induced distortions can be corrected (FSL topup).

Preprocessing¶

DELTA-SVD expects preprocessed input, so apply a state-of-the-art pipeline before running it:

• Visual QC of the raw data — inspect for artefacts and gross problems before preprocessing.
• Conversion — convert from DICOM with a tool that preserves the metadata required for preprocessing (phase-encoding direction, total readout time, gradient table), for example dcm2niix with BIDS sidecars. If you intend to continue with MRtrix3, converting with mrconvert to the .mif format is a convenient alternative, as it embeds this metadata directly in the image file.
• Core steps — denoising (e.g. MP-PCA, via MRtrix3's dwidenoise) and Gibbs-ringing removal (MRtrix3's mrdegibbs), followed by susceptibility-distortion correction (FSL topup) and eddy-current/motion correction (FSL eddy).
• Brain mask — produce a DWI-space brain mask; DELTA-SVD takes it as input.