Anatomical processing

qsiprep.workflows.anatomical.volume.init_anat_preproc_wf()

QSIPrep uses the anatomical reference for three things: a brain mask, a tissue segmentation that the reports draw over the distortion correction results, and the definition of the subject’s ACPC space that every output is written in. It also normalizes the reference to a template, so that atlases can be brought into the subject’s space later.

The reference can be the T1w or the T2w (--anat-modality). Whichever is chosen, the steps are:

1. Conform. Every image of the reference contrast is reoriented to LPS+ and resampled to a common voxel size.

2. Merge. If there are several images, they are N4-corrected and aligned to one another. With --subject-anatomical-reference first-lex (the default) they are registered to the first image in lexical order; with unbiased an unbiased mid-point template is built instead. The extra cost of the unbiased template is small for two images and about an order of magnitude for three or more.

3. Brain extraction with SynthStrip [1].

4. Tissue segmentation with SynthSeg [2], giving the tissue-class dseg and the regional desc-aseg_dseg derivatives.

   

   The SynthSeg tissue classes drawn over the anatomical reference, as shown in the report.

   If the other contrast is present too (a T2w alongside a T1w reference), it is registered to the reference with an affine antsRegistration, initialized by an antsAI rotation search, and written as an additional derivative.

5. AC-PC alignment. The reference is rigidly registered to the template, and the rigid transform defines ACPC space: the origin at the anterior commissure and the axes aligned with the template’s, but with the subject’s own brain shape and size. All outputs are written in this space.

6. Spatial normalization. antsRegistration registers the reference to the template in a multi-scale, mutual-information based, nonlinear scheme. The forward and inverse transforms are written with the anatomical derivatives. --skip-anat-based-spatial-normalization skips this step.

   The subject’s brain mask restricts this registration. The default SynthStrip mask keeps CSF and dura at the brain border, whereas the template’s brain mask is tight around the brain, so in atrophied brains the registration can pull the dura, rather than the cortex, onto the template’s brain edge. --force no-csf-synthstrip runs a second SynthStrip with its --no-csf model and uses that tighter mask for the nonlinear registration only. The AC-PC alignment, the saved desc-brain_mask images and the DWI workflows keep the default mask, so the flag changes the from-ACPC_to-<template> transforms and nothing downstream of the masks.

   The report’s normalization figure alternates between the template and the normalized T1w while the pointer is over it.

The preprocessed reference defines ACPC space. With several input images, that space is not exactly aligned with any one of them; the from-orig_to-anat transforms map each input into it.

(Source code, png, svg, pdf)

Lesions

For patients with focal lesions (stroke, tumor resection), a binary lesion mask (1 inside the lesion) in the same space and resolution as the T1w, named sub-<label>_label-lesion_roi.nii.gz in anat/, is used as a mask during spatial normalization [3], so that healthy tissue is not warped into the lesion or the other way around.

Sessions

By default the anatomical images of all sessions are merged into one reference and one ACPC space per subject. --subject-anatomical-reference sessionwise processes each session with its own reference, its own space and its own report, which is required with --infant and appropriate whenever the anatomy changes between sessions.

Infant data

--infant replaces the adult template with the MNIInfant cohort matching the participant’s age (see Running QSIPrep for where the age is read from). The cohort selection follows Nibabies. Brain extraction and segmentation are the same as for adults.

No anatomical image

With --anat-modality none there is no anatomical workflow. The b=0 reference is registered directly to the template and only the rigid part of that registration is kept, which gives an AC-PC aligned b=0 that keeps the shape and size of the original. There is no tissue segmentation in the reports and no template normalization.

References

[1]

Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann. Synthstrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. doi:10.1016/j.neuroimage.2022.119474.

[2]

Benjamin Billot, Douglas N. Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V. Dalca, and Juan Eugenio Iglesias. Synthseg: segmentation of brain mri scans of any contrast and resolution without retraining. Medical Image Analysis, 86:102789, 2023. doi:10.1016/j.media.2023.102789.

[3]

Matthew Brett, Alexander P. Leff, Chris Rorden, and John Ashburner. Spatial normalization of brain images with focal lesions using cost function masking. NeuroImage, 14(2):486–500, 2001. doi:10.1006/nimg.2001.0845.