Analysis

Pipeline outline

• Decide the resolution and orientation at which to do the registration
  • To keep the original image orientation ( downsampleGauss )
  • To rotate by 45 degrees to bring into "canonical orientation" ( canonicalDownsample )
• Decide whether to include mirror images of the subjects ( flipDownsample )
• Build an initial starting point, and pad if desired ( paddedAverage )
• Run the groupwise registration ( antsBuildTemplateGroupwise )
• Render at the original resolution ( runRenderHiRes )
• Average ( averageTally )

Methods for template building

• ANTs groupwise (antsBuildTemplateGroupwise)
• ANTs single (singleSubjectTemplateParallel.sh)
• CMTK groupwise (cmtkGroupwiseTemplate)
• CMTK single (cmtkSingleTemplate)

Analysis

• Estimate registration variance (accuracy) by comparing flips ( varianceLR )
• Estimate a "symmetrizing" transformation ( runSymmetry )
• Get a tranformation to "canonical" orientation - where the y-axis is the axis of symmetry ( sym2Canonical )
• Fit a Rayleigh distribution pixel- (or local window)-wise across subjects (perWindowRayleigh)

pipeline script

The templateAnalysis script runs many of the above analysis steps in sequence:

Usage:
  templateAnalysis [OPTIONS]
  -p [PREFIX] Specify the prefix for output of template building (Default='ALLF')
  -t [TEMPLATE] Specifies the template to use for analyis (Default='$PREFIX-template.nii.gz')
  -s Compute a symmetrizing transformation
  -v Compute flip-variance
  -r Estimate pixelwise Rayleigh distribution of flip-Variance
  -m Estimate pixelwise mean and variance of flip-Variance
  -c Estimate pixelwise percentiles distribution of flip-Variance (for percentiles = {0, 10, 50, 90, 100})

Skeleton analysis script

The evalAnalysis script works on a directory in which neuron and skeleton images have been transformed into template space, and measures proxies of registration accuracy (i.e. skeleton distance, and neuron image similarity)

Usage:
  evalAnalysis [OPTIONS]
  -p [PREFIX] Specify the prefix for output of template building (Default='ALLF')
  -t [TEMPLATE] Specifies the template to use for analyis (Default='$PREFIX-template.nii.gz')
  -e [RESOLUTION] Specifies the resolution at which to perform distance transform analysis
  -c Compartment label file
  -d Do the distance transform analysis
  -l Do the distance transform analysis per label
  -j Compute jacobian determinant maps for transforms
  -n Transform neuron images
  -r Generate report page
  -s Do neuron similarity
  -v Visualize neuron overlay

Examples

Finding a "symmetrizing" transformation

runSymmetry takes one argument - the image to be symmetrized. Outputs a flipped image (about the line y=x :this makes sense for the 63x drosophila image data ) as well as two transformations : one that produces the flip, and one that registers the flip to the original. The concatenation of these two transforms is the "symmetrizing" transformation

runSymmetry <the-image>

Estimating registration accuracy via flips

varianceLR <output image of distance> \
  <template> <flipped-template> <affine-that-flips-the-template-LR> \
  <subject-affine> <subject-deformation> \
  <flipped-subject-affine> <flipped-subject-deformation> \
  <affine-that-flips-the-subject>