Preprocessing

Building a pipeline

QSIPrep builds a pipeline based on your BIDS inputs. In general the pipeline will incorporate all the data it knows how to handle (i.e. fieldmaps, dMRI and anatomical data) automatically. There may be cases where you want to change the default behavior, particularly in regard to

Merging multiple scans from a session

For q-space imaging sequences it is common to have multiple separate scans to acquire the entire sampling scheme. These scans get aligned and merged into a single DWI series before reconstruction. It is also common to collect a DWI scan (or scans) in the reverse phase encoding direction to use for susceptibility distortion correction (SDC).

This creates a number of possible scenarios for preprocessing your DWIs. These scenarios can be controlled by the --separate_all_dwis argument. If your study has multiple sessions, DWI scans will never be combined across sessions. Merging only occurs within a session.

If --separate-all-dwis is present in the commandline call, each dwi scan in the dwi directories will be processed independently. You will have one preprocessed output per each DWI file in your input.

Otherwise (default) the DWI scans will be merged (i.e. their images will be concatenated). The merging affects the pipeline at different stages. If all DWIs in a session are in the same PE direction, they will be merged into a single series. If there are two PE directions detected in the DWI scans and 'fieldmaps' is not in ignore, images are combined according to their PE direction, and their b0 reference images are used to perform SDC. Further complicating this is the FSL workflow, which combines distortion correction with eddy/motion correction and will merge scans with different PE directions.

If you have some scans you want to combine and others you want to preprocess separately, you can call qsiprep more than once with BIDS filters to process the different scans.

Using BIDS filters

BIDS filters allow users to filter the set of images available to QSIPrep at run time. BIDS filters should be stored in a json file and passed to QSIPrep with the --bids-filter-file option. Filters modify “queries”, which are used to find data for each data type. NOTE: this is illustrating how modalities are queried in general, and is not the format of the file you will send to --bids-filter-file. The queries in QSIPrep are:

{
    "fmap": {"datatype": "fmap"},
    "sbref": {"datatype": "func", "suffix": "sbref"},
    "flair": {"datatype": "anat", "suffix": "FLAIR"},
    "t2w": {"datatype": "anat", "suffix": "T2w"},
    "t1w": {"datatype": "anat", "suffix": "T1w"},
    "roi": {"datatype": "anat", "suffix": "roi"},
    "dwi": {"datatype": "dwi", "suffix": "dwi"}
}

Each query has several “entities”, which can be modified by filters. The list of supported entities is here. To filter data, modify the queries by changing one or more of the supported entities in the BIDS filter file. The general format of the filter file is:

{
  "query": { "entity": "value" }
}

The entities specified in the filter file are added to the queries, so you only need to include entities you want to use for filtering. For example, this could be the contents of a valid BIDS filter file:

{
    "t1w": { "session": "MR1" },
    "dwi": { "session": "MR1", "run": "1" }
}

this modifies the “t1w” and “dwi” queries, and filters both T1w and DWI scans to select session “MR1”. It also filters on the run number for DWI scans only.

Multiple runs can be selected by passing arrays. For example:

{
    "dwi": { "run": [2,3] }
}

filters the “dwi” query for runs 2 and 3.

You can enable regular expressions for more detailed filtering, for example:

{
    "t1w": { "acquisition": "(?i)mprage", "regex_search": "true" },
}

will do a case-insensitive match of “mprage” within the “t1w” query.

Denoising and Merging Images

The user can decide whether to do certain preprocessing steps and, if so, whether they are performed before or after the DWI series are concatenated. Specifically, image denoising (using dwidenoise or patch2self) can be disabled with --denoise-method none. Gibbs unringing (using mrdegibbs) is disabled by default but can be enabled with --unringing-method mrdegibbs. B1 bias field correction is applied by default (using dwibiascorrect) and can be disabled with the --dwi-no-biascorr option. The intensity of b=0 images is harmonized across scans (i.e. scaled to an average value) by default, but this can be turned off using --dwi-no-b0-harmonization.

Together, denoising (MP-PCA or patch2self), Gibbs unringing B1 bias field correction and b=0 intensity normalization are referred to as denoising in QSIPrep. Each of these image processing operations has assumptions about its inputs and changes the distribution of noise in its outputs. Although the inclusion of each operation can be decided by the user, the order in which they are applied relative to one another is fixed. MP-PCA or patch2self are applied directly to the BIDS inputs, which should be uninterpolated and as “raw” as possible. Although Gibbs unringing should be performed on “raw” data, it is recommended in the MRtrix3 documentation to apply MP-PCA before Gibbs unringing. B1 bias field correction and b=0 intensity harmonization do not have as specific requirements about their inputs so are run last.

The last, and potentially very important decision, is whether the denoising operations are applied to each input DWI series individually or whether the denoising operations are applied to the concatenated input DWI files. At present, there is little data to guide this choice. The more volumes available, the more data MP-PCA/patch2self have to work with. However, if there if the head is in a vastly different location in different scans, denoising might be impacted in unpredictable ways.

Consider MP-PCA. If a voxel contains CSF in one DWI series and the subject repositions their head between scans so that the voxel contains corpus callosum in the next DWI series, the non-noise signal will be very different in the two series. Similarly, if the head is repositioned different areas will be closer to the head coil and therefore be inconsistently affected by B1 bias field. Similar problems can also occur within a DWI series due to subject head motion, but these methods have been shown to work well even in the presence of within-scan head movement. If the head position changes across scans is of a similar magnitude to that of within-scan head motion, it is likely fine to use the --denoise-after-combining option. To gauge how much between-scan motion occurred, users can inspect the Quality Control data to see whether Framewise Displacement is large where a new series begins.

By default, the scans in the same warped space are individually denoised before they are concatenated. When warped groups are concatenated an additional b=0 image intensity normalization is performed.

Preprocessing HCP-style

QSIPrep can be configured to produce a very similar pipeline to the HCP dMRI pipelines. HCP and HCP-Lifespan scans acquire complete multi-shell sequences in opposing phase encoding directions, making them a special case where Phase Encoding POLARity (PEPOLAR) techniques are used and the corrected images from both PE directions are averaged at the end. To produce output from qsiprep that is directly comparable to the HCP dMRI pipeline you will want to include:

--distortion-group-merge average \
--combine-all-dwis \

If you want to disable the image pair averaging and get a result with twice as many images, you can substitute average with concat.

Outputs of qsiprep

qsiprep generates three broad classes of outcomes:

  1. Visual QA (quality assessment) reports: one HTML per subject, depicting images that provide a sanity check for each step of the pipeline.

  2. Pre-processed imaging data such as anatomical segmentations, realigned and resampled diffusion weighted images and the corresponding corrected gradient files in FSL and MRTrix format.

  3. Additional data for subsequent analysis, for instance the transformations between different spaces or the estimated head motion and model fit quality calculated during model-based head motion correction.

  4. Quantitative QA: A single-line csv file per subject summarizing subject motion, coregistration quality and image quality.

Visual Reports

qsiprep outputs summary reports, written to <output dir>/qsiprep/sub-<subject_label>.html. These reports provide a quick way to make visual inspection of the results easy. One useful graphic is the animation of the q-space sampling scheme before and after the pipeline. Here is a sampling scheme from a DSI scan:

_images/sampling_scheme.gif

A Q5 DSI sampling scheme before (left) and after (right) preprocessing. This is useful to confirm that the gradients have indeed been rotated and that head motion correction has not disrupted the scheme extensively.

Preprocessed data (qsiprep derivatives)

There are additional files, called “Derivatives”, written to <output dir>/qsiprep/sub-<subject_label>/.

Derivatives related to T1w files are nearly identical to those produced by FMRIPREP and can be found in the anat subfolder:

  • *T1w_brainmask.nii.gz Brain mask derived using ANTs’ antsBrainExtraction.sh.

  • *T1w_class-CSF_probtissue.nii.gz

  • *T1w_class-GM_probtissue.nii.gz

  • *T1w_class-WM_probtissue.nii.gz tissue-probability maps.

  • *T1w_dtissue.nii.gz Tissue class map derived using FAST.

  • *T1w_preproc.nii.gz Bias field corrected T1w file, using ANTS’ N4BiasFieldCorrection

  • *T1w_space-MNI152NLin2009cAsym_brainmask.nii.gz Same as _brainmask above, but in MNI space.

  • *T1w_space-MNI152NLin2009cAsym_class-CSF_probtissue.nii.gz

  • *T1w_space-MNI152NLin2009cAsym_class-GM_probtissue.nii.gz

  • *T1w_space-MNI152NLin2009cAsym_class-WM_probtissue.nii.gz Probability tissue maps, transformed into MNI space

  • *T1w_space-MNI152NLin2009cAsym_dtissue.nii.gz Same as _dtissue above, but in MNI space

  • *T1w_space-MNI152NLin2009cAsym_preproc.nii.gz Same as _preproc above, but in MNI space

  • *T1w_space-MNI152NLin2009cAsym_target-T1w_warp.h5 Composite (warp and affine) transform to map from MNI to T1 space

  • *T1w_target-MNI152NLin2009cAsym_warp.h5 Composite (warp and affine) transform to transform T1w into MNI space

Derivatives related to diffusion images are in the dwi subfolder.

  • *_confounds.tsv A tab-separated value file with one column per calculated confound and one row per timepoint/volume

Volumetric output spaces include T1w (default) and MNI152NLin2009cAsym.

  • *dwiref.nii.gz The b0 template

  • *desc-brain_mask.nii.gz The generous brain mask that should be reduced probably

  • *desc-preproc_dwi.nii.gz Resampled DWI series including all b0 images.

  • *desc-preproc_dwi.bval, *desc-preproc_dwi.bvec FSL-style bvals and bvecs files. These will be incorrectly interpreted by MRTrix, but will work with DSI Studio and Dipy. Use the .b file for MRTrix.

  • desc-preproc_dwi.b The gradient table to import data into MRTrix. This and the _dwi.nii.gz can be converted directly to a .mif file using the mrconvert -grad _dwi.b command.

  • *bvecs.nii.gz Each voxel contains a gradient table that has been adjusted for local rotations introduced by spatial warping.

  • *cnr.nii.gz Each voxel contains a contrast-to-noise model defined as the variance of the signal model divided by the variance of the error of the signal model.

Confounds

See implementation on init_dwi_confs_wf.

For each DWI processed by qsiprep, a <output_folder>/qsiprep/sub-<sub_id>/func/sub-<sub_id>_task-<task_id>_run-<run_id>_confounds.tsv file will be generated. These are TSV tables, which look like the example below:

framewise_displacement        trans_x trans_y trans_z rot_x   rot_y   rot_z   hmc_r2  hmc_xcorr       original_file   grad_x  grad_y  grad_z  bval

n/a    -0.705 -0.002  0.133   0.119   0.350   0.711   0.941   0.943   sub-abcd_dwi.nii.gz     0.000   0.000   0.000   0.000
16.343        -0.711  -0.075  0.220   0.067   0.405   0.495   0.945   0.946   sub-abcd_dwi.nii.gz     0.000   0.000   0.000   0.000
35.173        -0.672  -0.415  0.725   0.004   0.468   1.055   0.756   0.766   sub-abcd_dwi.nii.gz     -0.356  0.656   0.665   3000.000
45.131        0.021   -0.498  1.046   0.403   0.331   1.400   0.771   0.778   sub-abcd_dwi.nii.gz     -0.935  0.272   0.229   3000.000
37.506        -0.184  0.117   0.723   0.305   0.138   0.964   0.895   0.896   sub-abcd_dwi.nii.gz     -0.187  -0.957  -0.223  2000.000
16.388        -0.447  0.020   0.847   0.217   0.129   0.743   0.792   0.800   sub-abcd_dwi.nii.gz     -0.111  -0.119  0.987   3000.000

The motion parameters come from the model-based head motion estimation workflow. The hmc_r2 and hmc_xcorr are whole-brain r^2 values and cross correlation scores (using the ANTs definition) between the model-generated target image and the motion-corrected empirical image. The final columns are not really confounds, but book-keeping information that reminds us which 4d DWI series the image originally came from and what gradient direction (grad_x, grad_y, grad_z) and gradient strength bval the image came from. This can be useful for tracking down mostly-corrupted scans and can indicate if the head motion model isn’t working on specific gradient strengths or directions.

Quality Control data

A single-line csv file (desc-ImageQC_dwi.csv) is created for each output image. This file is particularly useful for comparing the relative quality across subjects before deciding who to include in a group analysis. The columns in this file come from DSI Studio’s QC calculation and is described in [Yeh2019]. Columns prefixed by raw_ reflect QC measurements from the data before preprocessing. Columns prefixed by t1_ or mni_ contain QC metrics calculated on the preprocessed data. Motion parameter summaries are also provided, such as the mean and max of framewise displacement (mean_fd, max_fd). The max and mean absolute values for translation and rotation are max_translation and max_rotation and the maxima of their derivatives are in max_rel_translation and max_rel_rotation. Finally, the difference in spatial overlap between the anatomical mask and the anatomical brain mask and the DWI brain mask is calculated using the Dice distance in t1_dice_distance and mni_dice_distance.

Confounds and “carpet”-plot on the visual reports

fMRI has been using a “carpet” visualization of the BOLD time-series (see [Power2016]), but this type of plot does not make sense for DWI data. Instead, we plot the cross-correlation value between each raw slice and the HMC model signal resampled into that slice. This plot is included for each run within the corresponding visual report. Examples of these plots follow:

_images/sub-abcd_carpetplot.svg

For SHORELine higher scores appear more yellow, while lower scores are more blue. Not all slices contain the same number of voxels, so the number of voxels in the slice is represented in the color bar to the left of the image plot. The more yellow the pixel, the more voxels are present in the slice. Purple pixels reflect slices with fewer brain voxels.

_images/sub-pnc_carpetplot.png

For eddy slices with more outliers appear more yellow, while fewer outliers is more blue.

Preprocessing pipeline details

qsiprep adapts its pipeline depending on what data and metadata are available and are used as the input.

Processing the Subject Anatomical Reference T1w or T2w images

qsiprep.workflows.anatomical.init_anat_preproc_wf

(Source code)

As of version 0.18 QSIPrep has been changed to be very flexible with anatomical processing workflows. Versions prior to 0.18 were focused on the T1w images and provided only 2 possible templates. Version 0.18 introduces 2 terms that simplify the anatomical processing and open up new opportunities for choosing a template. First, is the subject anatomical reference and the second is the template anatomical reference.

As a dMRI-focused tool, QSIPrep only uses an anatomical reference image for an extra-robust brain extraction and to get a tissue segmentation for visualizing the susceptibility distortion correction results. The anatomical worflows leverage fast and powerful tools from FreeSurfer, namely SynthStrip and SynthSeg to perform brain extraction and segmentation.

Many imaging protocols acquire some high-resolution, undistorted anatomical reference scans. QSIPrep can use either T1-weighted ot T2-weighted 3D images as the anatomical reference. To specify which contrast you’d like to use for your anatomical reference, be sure to specify --anatomical-contrast as either T1w, T2w or none. Specifying none is equivalent to the previous option of --dwi-only, where no anatomical images are used from the input data and the AC-PC alignment is based either on the adult or infant MNI templates.

We discourage the use of --anatomical-contrast none in most cases. It is very rare to have dMRI data without any kind of T1w or T2w image from the same individual.

Regardless of whether you are using T1w or T2w images as your anatomical reference, the following steps will be applied to the anatomical reference images: Processing the Anatomical Reference images

  1. The anatomical sub-workflow begins by constructing an average image by conforming all found T1w or T2w images to LPS+ orientation and a common voxel size.

  2. If there are multiple images of the preferred anatomical contrast, they will be bias corrected using N4 and aligned to one another. If --longitudinal is specified they will be unbiasedly registered to each other using ANTs. Otherwise all the images are registered to the first image (see Longitudinal T1w processing).

  3. Brain extraction is performed using SynthStrip.

_static/brainextraction_t1.svg:scale:100%
_static/segmentation.svg:scale:100%

Handling Lesions and abnormalities

When processing images from patients with focal brain lesions (e.g. stroke, tumor resection), it is possible to provide a lesion mask to be used during spatial normalization to MNI-space [Brett2001]. ANTs will use this mask to minimize warping of healthy tissue into damaged areas (or vice-versa). Lesion masks should be binary NIfTI images (damaged areas = 1, everywhere else = 0) in the same space and resolution as the T1 image, and follow the naming convention specified in BIDS Extension Proposal 3: Common Derivatives (e.g. sub-001_T1w_label-lesion_roi.nii.gz). This file should be placed in the sub-*/anat directory of the BIDS dataset to be run through qsiprep.

_images/T1MNINormalization.svg

Animation showing T1w to MNI normalization

Longitudinal T1w processing

In the case of multiple T1w images (across sessions and/or within a session), T1w images are merged into a single template image using FreeSurfer’s mri_robust_template. This template may be unbiased, or equidistant from all source images, or aligned to the first image (determined lexicographically by session label). For two images, the additional cost of estimating an unbiased template is trivial and is the default behavior, but, for greater than two images, the cost can be a slowdown of an order of magnitude. Therefore, in the case of three or more images, qsiprep constructs templates aligned to the first image, unless passed the --longitudinal flag, which forces the estimation of an unbiased template.

Note

The preprocessed T1w image defines the T1w space. In the case of multiple T1w images, this space may not be precisely aligned with any of the original images. Reconstructed surfaces and functional datasets will be registered to the T1w space, and not to the input images.

Processing Infant Data

When processing infant DWI data, users may add --infant to their QSIPrep call. This will swap the default MNI152NLin2009cAsym template with the MNI infant template. It is highly advisable to also include --dwi-only to avoid problems with T1w skull-stripping.

DWI preprocessing

qsiprep.workflows.dwi.base.init_dwi_preproc_wf

(Source code)

Preprocessing of DWI files is split into multiple sub-workflows described below.

Head-motion / Eddy Current/ Distortion correction (FSL)

qsiprep.workflows.dwi.fsl.init_fsl_hmc_wf

FSL provides the most widely-used tools for head motion correction, eddy current correction and susceptibility distortion correction. These tools are designed to work directly with one another and share a file format that is unique to their workflow.

To ensure that the FSL workflow works as intended, all inputs are forced into to the FSL standard orientation. The head motion, eddy current and suscebtibility distortion corrections are applied at the end of eddy, which means that there will be two total interpolations in the FSL-based qsiprep workflow, as the final interpolation into T1w/AC-PC space is done externally in ANTs.

The FSL workflow can take three different forms.

  1. No distortion correction

  2. PEPOLAR distortion correction (using topup)

  3. Fieldmap-based distortion correction

No distortion correction

If there are no fieldmap images or the user has specified --ignore fieldmaps, no distortion correction will occur. In this case, only head motion correction and eddy current correction will be performed. The workflow looks like this:

(Source code)

PEPOLAR (TOPUP) Distortion Correction

When images with different phase encoding directions are available, either dedicated fieldmaps (in the fmap/ directory) or DWI series (in the dwi/ directory), example b=0 images can be

(Source code)

Fieldmap-based Distortion Correction

If a GRE fieldmap or SyN-based fieldmapless distortion correction are detected, these will be performed on the outputs of eddy. For details see Susceptibility correction methods.

(Source code)

Configuring eddy

eddy has many configuration options. Instead of making these commandline options, you can specify them in a JSON file and pass that to qsiprep using the --eddy-config option. An example (default) eddy config json can be viewed or downloaded here

Head-motion estimation (SHORELine)

qsiprep.workflows.dwi.hmc.init_dwi_hmc_wf

A long-standing issue for q-space imaging techniques, particularly DSI, has been the lack of motion correction methods. DTI and multi-shell HARDI have had eddy_correct and eddy in FSL, but DSI has relied on aligning the interleaved b0 images and applying the transforms to nearby non-b0 images.

qsiprep introduces a method for head motion correction that iteratively creates target images based on 3dSHORE or MAPMRI fits. First, all b0 images are aligned to a midpoint b0 image (or the first b0 image if hmc_align_to="first") and each non-b0 image is transformed along with its nearest b0 image.

Then, for each non-b0 image, a 3dSHORE or MAPMRI model is fit to all the other images with that image left out. The model is then used to generate a target signal image for the gradient direction and magnitude (i.e. q-space coordinate) of the left-out image. The left-out image is registered to the generated target signal image and its vector is rotated accordingly. A new model is fit on the transformed images and their rotated vectors. The leave-one-out procedure is then repeated on this updated DWI and gradient set.

If "none" is specified as the hmc_model, then only the b0 images are used and the non-b0 images are transformed based on their nearest b0 image. This is probably not a great idea.

Susceptibility distortion correction is run as part of this pipeline to be consistent with the TOPUP/eddy workflow.

Ultimately a list of 6 (or 12)-parameters per time-step is written and fed to the confounds workflow. These are used to estimate framewise displacement. Additionally, measures of model fits are saved for each slice for display in a carpet plot-like thing.

(Source code)

Susceptibility correction methods

qsiprep.workflows.fieldmap.base.init_sdc_wf

_images/unwarping.svg

The are three kinds of SDC available in qsiprep:

  1. Phase Encoding POLARity (PEPOLAR) techniques (also called blip-up/blip-down): This is the implementation from sdcflows, using 3dQwarp to correct a DWI series using a fieldmap in the fmaps directory [Jezzard1995]. The reverse phase encoding direction scan can come from the fieldmaps directory or the dwi directory. If using Head-motion / Eddy Current/ Distortion correction (FSL), then TOPUP is used for this correction. Also relevant is Selecting representative images for PEPOLAR.

  2. Phase-difference B0 estimation: Use a B0map sequence that includes at lease one magnitude image and two phase images or a phasediff image.

  3. Fieldmap-less estimation (experimental): The SyN-based susceptibility distortion correction implemented in FMRIPREP. To use this method, include argument --use-syn-sdc when calling qsiprep. Briefly, this method estimates a SDC warp using ANTS SyN based on an average fieldmap in MNI space. For details on this method.

qsiprep determines if a fieldmap should be used based on the "IntendedFor" fields in the JSON sidecars in the fmap/ directory.

Selecting representative images for PEPOLAR

TOPUP estimates EPI distortion based on the shapes of images with different phase encoding directions and total readout times (i.e. warped groups). It is therefore ideal to provide less-noisy images as inputs, so the registration has plenty of accurate anatomical features to work with.

For diffusion-weighted MRI, the b=0 images are used as input to TOPUP. While these contain a lot of anatomical detail, they can also contain troublesome artefacts such as spin history, head motion and slice dropout.

In QSIPrep versions up until 0.13, up to 3 b=0 images were selected per warped group as input to TOPUP. The images were selected to be evenly spaced within their acquisitions.

In versions 0.13 and later, QSIPrep finds the “most representative” b=0 images per warped group. A nearly identical approach is used in the developmental HCP pipelines, where a pairwise spatial correlation score is calculated between all b=0 images of the same warped group and the images with the highest average correlation to the other images are used as input to TOPUP. To see which images were selected, examine the selected_for_topup column in the confounds tsv file.

Using only DWI data (bypassing the T1w workflows)

It is possible to use QSIPrep to process only diffusion-weighted images. In the case of infant data, where robust skull-stripping methods are not currently available, or where anatomical preprocessing has already been performed in another pipeline, the user can specify --dwi-only.

Instead of registering the b=0 template image to the skull-stripped T1w image, the b=0 template is registered directly to a template and only the rigid part of the transformation is kept. This results in an AC-PC aligned b=0 template that maintains the shape and size of the original image.

In this case the b0_anat_coreg workflow instead registers the b=0 reference to an AC-PC-oriented template and the rigid components of the coregistration transform are extracted.

DWI reference image estimation

qsiprep.workflows.dwi.util.init_dwi_reference_wf

(Source code)

This workflow estimates a reference image for a DWI series. This procedure is different from the DWI reference image workflow in the sense that true brain masking isn’t usually done until later in the pipeline for DWIs

Pre-processed DWIs in a different space

qsiprep.workflows.dwi.resampling.init_dwi_trans_wf

_images/preprocessing-8.png

(Source code, png, svg, pdf)

A DWI series is resampled to an output space. The output_resolution is specified on the commandline call. All transformations, including head motion correction, susceptibility distortion correction, coregistration and (optionally) normalization to the template is performed in a single shot using a Lanczos kernel.

There are two ways that the gradient vectors can be saved. This workflow always produces a FSL-style bval/bvec pair for the image and a MRTrix .b gradient table with the rotations from the linear transforms applied. You can also write out a local_bvecs file that contains a 3d vector that has been rotated to account for nonlinear transforms in each voxel. I’m not aware of any software that can use these yet, but it’s an interesting idea.

b0 to T1w registration

qsiprep.workflows.dwi.registration.init_b0_to_anat_registration_wf

_images/preprocessing-9.png

(Source code, png, svg, pdf)

This just uses antsRegistration.