DWI masking

Many DWI processing operations either necessitate the use of a binary mask image in order to spatially constrain the operation in some way, or can be executed in less time by not performing the relevant calculations in those voxels that are not of interest. However, while this would seem like a relatively trivial operation, it is in fact deceptively difficult to devise an appropriate heuristic for deriving an appropriate mask that works for a wide range of DWI data. It is not uncommon for the derivation of this mask to go awry in a range of scenarios, which can have serious implications for downstream processing steps. For this reason, as of MRtrix version 3.1.0, a range of DWI mask derivation algorithms are provided, allowing users to assess which heuristics work best for their particular data. The purpose of this documentation page is to describe those algorithms that are available, the circumstances in which they may or may not work, the features that are available for users to manipulate this behaviour, and the applications in which these details are most relevant for user attention.

dwi2mask algorithms

dwi2mask 3dautomask

Provides the mean b=0 image directly to AFNI command 3dAutomask .

dwi2mask ants

Provides the mean b=0 image directly to ANTs command antsBrainExtraction.sh, configured for T2-weighted image input.

This algorithm necessitates the specification of a template image and corresponding binary mask image defined on that template. These two images must be provided by the user either using the -template command-line option, or the Dwi2maskTemplateImage and Dwi2maskTemplateMask configuration file options (see Configuration file options).

dwi2mask b02template

Registers the subject’s mean b=0 image to a template image, and back-propagates a binary mask back into the individual’s DWI voxel grid. Achieved as follows:

  1. Non-linearly register subject’s mean b=0 image to a specified template image;

  2. (If not calculated implicitly as part of step 1) Invert the non-linear deformation field;

  3. Transform binary mask associated with template image onto voxel grid of mean b=0 image (with interpolation);

  4. Apply a threshold of 0.5 to transformed image to produce a mask.

There are multiple external software tools that can be utilised for performing the core image registration and transformation processes:

  • antsquick: Utilises the ANTs command antsRegistrationSyNQuick.sh for registration; transforms mask image to subject space using ANTs command antsApplyTransforms.

  • antsfull: Utilises the ANTs commands antsRegistration for registration, using the registration parameters specified in the article:

    Tustison, Nicholas J., and Brian B. Avants. Explicit B-Spline Regularization in Diffeomorphic Image Registration. Frontiers in Neuroinformatics 7 (December 23, 2013): 39. https://doi.org/10.3389/fninf.2013.00039.

    Template mask image is then transformed to subject space using ANTs command antsApplyTransforms.

  • fsl: Utilises FSL commands as follows:

    • flirt: Initial affine registration;

    • fnirt: Non-linear registration;

    • invwarp: Inversion of warp from subject to template;

    • applywarp: Transform template mask to subject space.

By default, if no manual selection is made here using either the -software command-line option or the Dwi2maskTemplateSoftware configuration file entry, the antsquick approach will be used.

This algorithm necessitates the specification of a template image and corresponding binary mask image defined on that template. These two images must be provided by the user either using the -template command-line option, or the Dwi2maskTemplateImage and Dwi2maskTemplateMask configuration file options (see Configuration file options).

The registration operation can be expected to perform best if the specified template image is of comparable shape and image contrast to that of the b=0 volumes of the DWI data being processed. As such, if using an existing template image, a T2-weighted image would be recommended. Alternatively, one could produce a population template b=0 image based on one’s own data, and manually define a mask on that template that could then subsequently be used for DWI masking.

For all registration algorithms, there are dwi2mask command-line options available for fine-tuning the behaviour of the registration by passing command-line options down to the relevant command(s); further, it is possible to set such parameters within the MRtrix configuration file, which may be of particular use if configuration file option Dwi2maskAlgorithm is set to b02template (see Configuration file options).

dwi2mask consensus

This algorithm is unique compared to all other dwi2mask algorithms, in that it does not provide one specific heuristic for DWI mask estimation; instead, it executes all other dwi2mask algorithms, and produces a single mask based on the consensus of those algorithms. Currently this consensus is simply those voxels that were included in the estimated masks of more than 50% of the algorithms utilised. Note that if the external software requirements of any specific dwi2mask algorithm are not installed, the consensus algorithm will report that not all algorithms could be executed, and will utilise only the outputs of those algorithms that could be executed successfully.

dwi2mask fslbet

Provides the mean b=0 image directly to FSL command bet.

dwi2mask hdbet

Provides the mean b=0 image directly to HD-BET command hd-bet.

dwi2mask legacy

Reproduces the behaviour of the dwi2mask binary executable that was included in MRtrix3 prior to version 3.1.0.

It involves the following steps:

  1. Compute the mean diffusion-weighted signal intensity for each b-value;

  2. For each b-value independently, automatically determine a threshold to apply to produce a binary mask;

  3. Sum the masks from step 2 across b-values;

  4. Apply a median filter;

  5. Select the largest connected component and fill holes;

  6. Apply mask cleaning filter to remove small areas only connected to the largest component via thin “bridges”.

dwi2mask mean

A heuristic algorithm that is based on simply taking the mean DWI intensity across all volumes, and then applying a threshold. It was reported to provide good results for some forms of data, but is not necessarily guaranteed to do so for other DWI acquisition protocols; algorithm dwi2mask trace is intended to operate on a similar concept, but be more robust against variations in acquisition.

Operations are as follows:

  1. Compute the mean DWI intensity across all volumes, regardless of b-value;

  2. Automatically determine an intensity threshold for this image to produce a binary mask;

  3. Select the largest connected component and fill any holes;

  4. Apply mask cleaning filter to remove small areas only connected to the largest component via thin “bridges”.

dwi2mask mtnorm

This algorithm implements a subset of the functionalities provided in the dwibiasnormmask script (described in further detail below). It is based on utilisation of the results generated by the mtnormalise command. The basic premise is that, following multi-shell multi-tissue CSD and appropriate response function bias correction / bias field cocrrection / intensity normalisation, an image consisting of the sum of all macroscopic tissue ODFs should be approximately 1.0 in brain voxels and 0.0 in non-brain voxels.

The order of operations is as follows:

  1. If not provided by the user, generate an initial brain mask using the default dwi2mask algorithm.

  2. Perform three-tissue response function estimation.

  3. Perform multi-shell multi-tissue CSD (with all three macroscopic tissues—WM, GM and CSF—if possible, otherwise only WM and CSF)

  4. Use mtnormalise to correct:

    1. Biases in response function magnitudes using -balanced option (note that this functionality is deliberately omitted from typical quantitative analysis pipelines as it may regress out effects of interest)

    2. Smoothly-varying bias field

    3. Global intensity scaling

  5. Calculate an image of the sum of tissue ODFs

  6. Apply a threshold to binarize this image (default threshold is 0.5).

  7. Apply mask cleaning operations (eg. largest connected component).

dwi2mask synthstrip

The SynthStrip method is based on a deep learning neural network that has been trained on a wide range of neuroimaging modalities and data qualities. This algorithm provides the mean b=0 image to SynthStrip, whether installed as part of FreeSurfer (version 7.3.0 or later) or as the stand-alone Singularity container.

dwi2mask trace

Heuristic algorithms for generating masks from DWI data based on trace-weighted images (i.e. mean image intensity within each shell) in a manner different to that of the dwi2mask legacy algorithm.

Its behaviour is as follows:

  1. Calculate the trace-weighted image for each shell;

  2. For each shell, find a multiplicative factor that gives the trace-weighted image approximately the same intensity of that of the first shell (this is so that each shell contributes approximately equally toward determination of the mask);

  3. Calculate the mean trace-weighted image across shells;

  4. Automatically determine an intensity threshold for this image to produce a binary mask;

  5. Select the largest connected component and fill any holes;

  6. Apply mask cleaning filter to remove small areas only connected to the largest component via thin “bridges”;

  7. If the command-line option -iterative is not used, the algorithm ceases at this point (i.e. the default behaviour);

  8. For each b-value shell, compute the mean and standard deviation of the trace-weighted image intensities inside and outside of the current mask, and use this to derive Cohen’s d statistic;

  9. Perform a recombination of the trace-weighted images; but the multiplicative weights applied to each b-value shell trace image are, instead of being based on intensity matching as in step 2, the Cohen’s d statistics calculated in step 8;

  10. Apply a threshold and mask filtering operations as in steps 4–6;

  11. If the resulting mask differs from the previous estimate, go back to step 8; if not, or if a maximum number of iterations is reached, the algorithm is completed.

Note that the iterative version of this algorithm can currently be considered a hypothetical heuristic, and it is not yet known whether or not its behaviour is reasonable across a range of DWI data; it should therefore be considered entirely experimental.

Algorithm comparison

Algorithm

External dependencies

Uses more that b=0

Assumptions

Robust to bias field

Can use GPU

3dAutomask

Yes (AFNI)

No

Unknown

Unknown

No

ants

Yes (ANTs)

No

Brain; WM darker than GM

Unknown

No

b02template

Yes (ANTs / FSL)

No

Matches template

Yes

No

consensus

Only if installed

Yes

Various

Various

No

fslbet

Yes (FSL)

No

Approx. spherical

Yes

No

hdbet

Yes (HD-BET)

No

Human brain

Yes

Yes

legacy

No

Yes

Single connected component

No

No

mtnorm

No

Yes

WM / GM / CSF constituency; single connected component

Yes

No

synthstrip

Yes

No

Human brain

Yes

Yes

trace

No

Yes

Single connected component

No

No

Python scripts utilising dwi2mask

There are a number of Python scripts provided within MRtrix3 that operate on DWI data and necessitate use of a mask, and therefore (if not provided with one explicitly at the command-line) will internally execute the dwi2mask command.

Because it is not possible for the user to manually specify how dwi2mask should be utilised in this scenario, there are configuration file options provided to assist in controlling the behaviour of dwi2mask in these scenarios (see below).

MRtrix3 Python command

Purpose of DWI mask

dwi2response

Voxels outside of the initial mask are never considered as candidates for response function(s), nor do they contribute to any optimisation of the selection of such.

dwibiascorrect

Only voxels within the mask are utilised in optimisation of bias field parameters.
For ants algorithm, field is estimated within the mask but applied to all voxels within the field of view (field basis is extrapolated beyond the extremities of the mask);
for fsl algorithm, field is both estimated within, and applied to, only those voxels within the mask, producing a discontinuity in image intensity at the outer edge of the mask that can be deleterious for subsequent quantitative analyses.

dwibiasnormmask

Determination of an initial brain mask by which to constrain the first iteration (see below).

dwifslpreproc

Constrains optimisation of distortion parameter estimates in FSL eddy.
If performing susceptibility distortion correction, this is applied to the DWI data subsequently to the appplication of FSL command applytopup.

dwigradcheck

Utilised as both seed and mask image for streamlines tractography in the tckgen command.

dwibiasnormmask

This new script is an experimental approach for improving DWI brain mask estimation (among other things), initially created during development of the MRtrix3_connectome BIDS App. It is based on the simple observation that the processes of bias field estimation, intensity normalisation, and brain mask derivation, can have circular dependencies between one another, and that therefore combining them into a single step may be beneficial. It is however noted that the behaviour of this algorithm can vary between different types of data, and therefore close scruitiny of such is recommended.

While this script is highly dependent on the operation of the mtnormalise command (as was observed to be the case for the dwi2mask mtnorm algorithm above, which performs a subset of the functionalities within dwibiasnormmask), the form of the primary results that it provides are slightly different:

  • Output intended for usage:

    With mtnormalise, a set of ODFs are provided as input, and a set of ODFs are then yielded as output, where the output ODFs have been corrected for a smoothly-spatially-varying bias field, and global intensity scaling (and importantly for quantitative applications the same intensity scaling is applied to all ODFs). For dwibiasnormmask, the provided input is a DWI series, and the yielded output is a DWI series, where the output has had the same smoothly-spatially-varying bias field and global intensity scaling corrections applied. The process of estimating these corrections is identical; the only difference is that in dwibiasnormmask the corrections are back-projected to correct the DWI series from which the ODFs were estimated, rather than directly utilising the corrected ODFs.

  • Global intensity normalisation:

    The topic of Global intensity normalisation is a long-standing issue in the domain of CSD analysis. Unlike other diffusion models, the b=0 intensity of each voxel is not used as a reference for the modelling of DWIs in that voxel, and even in the context of multi-tissue CSD the composition of the voxel is not explicitly forced to be unity. This does however raise the issue of how to appropriately globally scale the intensities of the image data in order for observed differences in eg. Apparent Fibre Density (AFD) to be attributable to the effect of interest rather then meaningless differential scaling of image intensities between subjects.

    The approach taken by mtnormalise is to determine the scaling factor that results in voxels throughout the brain having a sum of tissue densities of approximately unity. They will not all be exactly unity, even after bias field correction, but they should be approximately centred around unity.

    dwibiasnormmask provides a more advanced version of the original proposal for global intensity normalisation for AFD analysis. It was first proposed that the b=0 signal intensity in CSF should act as a reference intensity to normalise across subjects. However identifying appropriate exemplar voxels to do so can be labour-intensive and difficult. In dwibiasnormmask, information from the dwi2response dhollander response function estimation algorithm and the mtnormalise approach are combined in such a way that the b=0 CSF-like intensity is scaled to a fixed reference intensity:

    • using data from across the entire brain even in the presence of partial volume;

    • accounting for potential miscalibrations in response function estimation;

    • in conjunction with bias field estimation and correction.

    It is not yet known whether using this approach for global intensity normalisation may yield greater sensitivity to effects of interest. It should however be noted that if one were to subsequently execute mtnormalise and make use of its output ODFs, then the effective global intensity normalisation behaviour would revert to that of mtnormalise rather than that described above.

Another key aspect of this algorithm is the data used to derive the brain mask. Most DWI brain masking approaches base their operation on the mean b=0 image (see dwi2mask_algorithm_comparison above). In this algorithm, there is an alternative 3D image that can be used to drive brain mask derivation, being the sum of tissue ODFs. Depending on the configuration, this image may be used rather than the bias-field-corrected DWI series to estimate a new brain mask at each iteration; for instance, duplicating the functionality of the dwi2mask mtnorm algorithm above.

The script itself operates as follows:

  1. If no initial mask is provided, then one must be calculated using dwi2mask.

  2. Three-tissue response function estimation using dwi2response dhollander.

  3. Multi-shell multi-tissue CSD, by default using a lower WM lmax for computational efficiency.

  4. mtnormalise to estimate bias field and intensity scaling factors between tissues.

  5. Estimation of a new brain mask, using either the bias-field-corrected DWI series or the tissue ODF sum image.

  6. Determination of whether to exit, or loop back to step 2, based on:

    1. Adequate similarity of the brain mask between successive iterations;

    2. Masks between successive iterations becoming less similar rather than more similar (indicating some form of instability or divergence);

    3. Reaching maximal number of iterations.

While the primary output of the command is the DWI series corrected for bias field and intensity normalised, the brain mask corresponding to the last stable iteration can be additionally exported using the -output_mask option.

When initially developed, the number of iterations for this approach was fixed at 2, as the solution was found to erroneously diverge after that in some instances. It is however possible that for certain data, as well as the subsequent addition of the explicit check for mask divergence, it is possible to permit a larger number of iterations and allow the algorithm to converge toward a fixed solution. Feedback on the success or failure of this experimental script for different data is encouraged.

Configuration file options

There are many options that can be set within the MRtrix3 Configuration file that directly influence the operation of the dwi2mask command. These are included in the List of MRtrix3 configuration file options page, but are mentioned here also for discoverability:

  • Dwi2maskAlgorithm

    For those Python scripts utilising dwi2mask, this is the dwi2mask algorithm that will be invoked. If not explicitly set, the legacy algorithm will be used.

    Note

    Setting this configuration file option does not enable the utilisation of dwi2mask without manually specifying the algorithm to be used. For manual usage, the algorithm must always be specified. This option only controls the algorithm that will be used when dwi2mask is invoked from inside one of the Python scripts provided with MRtrix3.

  • Dwi2maskTemplateSoftware

    If dwi2mask b02template is invoked, and the -software command-line option is not used, the value of this option determines the software tool that will be utilised for registration to the template and back-propagation of the mask in template space to the subject’s DWI data. In the absence of this configuration file option, antsquick (i.e. ANTs antsRegistrationSyNQuick.sh) will be used.

  • Dwi2maskTemplateImage and Dwi2maskTemplateMask

    This pair of configuration file options allow the user to pre-specify the filesystem locations of the two images (T2-weighted template and corresponding binary mask) to be utilised by the dwi2mask ants and dwi2mask b02template algorithms. Note that there is no “default” template to be utilised by these algorithms; so the user must either include these entries in their configuration file, or manually specify the -template command-line option whenever they use dwi2mask ants or dwi2mask b02template. If the value of configuration file option “Dwi2maskAlgorithm” is “ants” or “b02template”, then these two entries must also be specified.

  • Dwi2maskTemplateANTsQuickOptions, Dwi2maskTemplateANTsFullOptions, Dwi2maskTemplateFSLFlirtOptions and Dwi2maskTemplateFSLFnirtConfig

    These options allow full automated control over the parameters with which the external neuroimaging software package registration commands are executed. If one of the relevant dwi2mask b02template command-line options is used explicitly (-ants_options, -flirt_options, -fnirt_config), that information takes precedence; otherwise, if one of these configuration file entries is set, that information will be propagated directly to the relevant command.