Imprecision in parcellations

[gkiar]

As discussed with @binarybottle, a potential project could explore the impact of uncertainty in anatomical labeling. This comparison could roughly be laid out as follows:

• Objective: characterize the impact of enforcing strict (and uncertain/imprecise) regional boundaries in neuroimaging analysis on reliability and predictive power. Optionally, check whether sampling + bootstrapping across repeated probabilistic labeling strategies can improve quality.

• Modalities: any of structural, functional, and/or diffusion

• Dataset would contain, if possible:

  1. Test-ReTest measurements,
  2. Reasonable sample size for a simple classification or regression analysis (S~100?)
  3. Manually annotated brain regions

• Processing: standard pipeline for the modalities, where features are ultimately downsampled prior to statistical analysis using one of the following parcellation schemes:

  1. Manual labels (if available)
  2. Automated probabilistic labeling (i.e. region determined by sampling a set of label probabilities)
  3. Automated deterministic labeling (i.e. a flat nifti file with region labels)
  4. Automated fuzzy labeling via deterministic labels (i.e. a flat nifti file with region labels, interpreted as a probability map within neighbouring voxels)

There are many open questions for the design, such as which parcellation(s) to use and which dataset(s) may exist with manual annotations already.

[gkiar]

From a conversation with Mike, his could be extended conceptually to include not just structural alignments to templates but functional alignments, as well (cc: @TingsterX)

[binarybottle]

Thank you for documenting our discussion points so faithfully, @gkiar -- Looking forward to working with you to advance these ideas!