Skip to content

1. Template

Jump to bottom
Ellyn Butler edited this page Mar 29, 2021 · 17 revisions

The final call to antsMultivariateTemplateConstruction2.sh is as follows:

antsMultivariateTemplateConstruction2.sh -d 3 -o /data/output/${projectName}Template_ -n 0 -i 5 -c 2 -j 8 -g .15 -m CC[2] -q 500x500x500x500x500x500x500x500x500x500x500x500x500x500x500x500x225x75x25,1e-6,10 -f 8x8x8x8x8x8x8x8x8x8x8x8x7x6x5x4x3x2x1 -s 8.057374066034377x7.6320909504467975x7.206734452718699x6.781290764178728x6.355742375591618x5.930066746813941x5.5042343571715975x5.078205771322212x4.651927085986457x4.225322606736783x3.798282560433022x3.370641399941647x2.942137020149432x2.5123277660148595x2.0804050381276458x1.6447045940431997x1.2011224087864498x0.735534255037358x0.0mm -z /data/output/MNI-1x1x1Head_pad.nii.gz /data/output/tmp_subjlist.csv

Steps of the template construction (Ellyn's understanding as of February 22, 2021):

  1. Rigidly align all of the SSTs in /data/output/tmp_subjlist.csv to the reference template specified with -z.
  2. Perform pairwise deformable registrations from each SST to the reference template, where each image is downsampled according to -f (shrinkage, or how much the registered image is allowed to change per iteration), the number of deformable stages is specified using -q, and the smoothing is specified using -s (the number supplied is the sigma for a multivariate Gaussian distribution). This is performed in stages, according to the numbers specified between the x's.
  3. Register the moving images (SSTs) pairwise to the template space and average them.
  4. Invert the average warp to the template space, and apply to the average of the registered SSTs. The inverse of the average warp is scaled in magnitude by -g, so lower values of -g will result is less change (i.e., smaller deformations) between template iterations. The output of this is the template for the next round of pairwise registrations.
  5. Repeat 2-4 -i times.

Details:
-n 0: No N4 bias field correction (was performed on the T1w images that comprise the SSTs)
-c 2: Allow for parallel computation using pexec
-j 8: Number of cpu cores (8) to use locally for pexec option
-g .15: Gradient step size. Changes the shape of the template at each iteration (-i). Lower values might help convergence, but will not improve pairwise registrations
-m CC[2]: Similarity metric is cross-correlation (CC), with a radius of 2, which defines the volume over which the cross-correlation is computed

Notes

  • Estimate how much the template is changing every iteration by calculating the Jacobian of template0warp.nii.gz
  • The average inverse warp is applied to try to change the shape of the template to better represent the population average. The best template has a mean transform close to 0 because the population transforms to the average point largely cancel out. It is also done to prevent small biases in pairwise registrations squishing the template over time.

http://www.ncbi.nlm.nih.gov/pubmed/20851191

http://www.ncbi.nlm.nih.gov/pubmed/19818860

Clone this wiki locally