Optimizing EPI to T1 MPRAGE registration

[nickcicero18]

Optimizing EPI to T1 MPRAGE registration

[gdevenyi]

Hi,

(Note the script and example data have brain.nii and brainmask.nii mixed up naming wise, I have preserved this in my example)

Here's my crack at optimizing the registration, changes made

• use the EPI brainmask to fully extract the EPI, so that there is "1:1 tissue correspondence" at least for the FOV of the scan
• BSpline interpolation because its much faster than Lancoz and very nearly as accurate (warning, this produces negative background values on resample you need to handle)
• don't winzorise the intensities, there's no huge outliers in either
• don't histogram match, they are not the same modality so I don't think this is appropriate (this is better for matching say, T1 FLASH to T1 MPRAGE)
• these scans should be "pre aligned" since they're the same subject same session, therefore don't do a COM --initial-moving-transform
• always use dense sampling and gradient filter: Mattes[ ${target},${moving},1,32,None,1,1 ]
• These are the same subject, use only Rigid transforms
• These are the same subject, don't use too much smoothing subsampling
• Do multiple stages of rigid, first aligning the whole-brain, and then refining alignment with the ROI mask
• Use the brainstem mask for the SyN registration

Note that at the end of this, the intention is to align JUST the brainstem, if I were doing a full registration where I want to explicitly also focus on the brainstem, I would do a SyN without the masks, and then refine the SyN in the masked area again.

This hard-coded script is based on the outputs from my generalized registration script antsRegistratioN_affine_SyN.sh from https://github.com/cobralab/minc-toolkit-extras/

P.S. The MPRAGE is a bit over-padded which will causes excess unneeded calculations that need to be done, cropping the FOV to be tighter would speed up the script.

#!/bin/bash

set -euo pipefail

set -x

target=brainmask.nii
moving=EPI_Tmean_extracted.nii.gz
output=ants_output  # Remove the extension if any.

target_mask_brainstem=brainstem_mask_anat.nii.gz
moving_mask_brainstem=brainstem_mask_EPI.nii.gz

ImageMath 3 EPI_Tmean_extracted.nii.gz m EPI_Tmean.nii.gz EPI_Tmean_mask.nii.gz

antsRegistration --dimensionality 3 --float 0 --verbose 1 \
        --output [$output,${output}.nii.gz] \
        --interpolation BSpline[5] \
        --transform Rigid[ 0.1 ] \
                --metric Mattes[ ${target},${moving},1,32,None,1,1 ] \
                --convergence [ 500x500x225x75,1e-6,10 ] \
                --shrink-factors 5x4x3x2 \
                --smoothing-sigmas 2.5x2.0x1.5x1.0mm \
                --masks [ NOMASK, NOMASK ] \
        --transform Rigid[ 0.1 ] \
                --metric Mattes[ ${target},${moving},1,64,None,1,1 ] \
                --convergence [ 225x75x25x25,1e-6,10 ] \
                --shrink-factors 3x2x1x1 \
                --smoothing-sigmas 1.5x1.0x0.5x0.0mm \
                --masks [ NOMASK, NOMASK ] \
        --transform Rigid[ 0.1 ] \
                --metric Mattes[ ${target},${moving},1,64,None,1,1 ] \
                --convergence [ 75x25x25,1e-7,10 ] \
                --shrink-factors 2x1x1 \
                --smoothing-sigmas 1.0x0.5x0.0mm \
                --masks [ ${target_mask_brainstem},${moving_mask_brainstem} ] \
          --transform SyN[ 0.1 ] \
                --metric CC[ ${target},${moving},1,4,None,1,1 ] \
                --convergence [ 500x180x60x20x20,1e-6,10 ] \
                --shrink-factors 4x3x2x1x1 \
                --smoothing-sigmas 2.0x1.5x1.0x0.5x0.0mm \
                --masks [ ${target_mask_brainstem},${moving_mask_brainstem} ]

Couldn't attach the file here, too large, but here's the ants_output files (resampled image, affine file, forward/reverse warp fields)

https://we.tl/t-EKohDy5p1j

[cookpa]

I ran the script as-is and it looks good to me. There is an element of randomness so results can vary, but I'm also using an up-to-date antsRegistration.

The bottom panel comes from running

antsApplyTransforms -d 3 -i EPI_Tmean.nii.gz -o ants_outputAffineOnlyWarped.nii.gz -r brainmask.nii \
  -t ants_output0GenericAffine.mat -n LanczosWindowedSinc

I always do this so I can see how well aligned the affine transform gets us. In this case it seems pretty close.

If it's failing to capture the nonlinear deformation there are a few things to try, in no particular order

• Boosting the SyN step size to 0.2

• Starting SyN at one level downsampled, ie -f 2x1

• BSplineSyN, this lets you smooth the update field with a spline in 3D, rather than an isotropic kernel. So you can allow more deformation along the AP axis. This can let you use a larger step size without having too much deformation in the other directions.

• Using landmark points in both images as an additional metric.

I also noticed that contrast range of the anatomical image is pretty heavily truncated. All the voxels in red here have constant intensity, and that's before the registration winsorization is applied

[gdevenyi]

I also noticed that contrast range of the anatomical image is pretty heavily truncated.

Metadata says the file is post-Freesurfer, so its had lots of manipulation done to it

[nickcicero18]

@cookpa @gdevenyi thank you both for your quick responses and insightful comments. I'll give these suggestions a try! Thank you again