title          :: Alignment (part 1/4): Background
presentation   :: afni14_alignment.pdf
scripts        ::
date           :: 2020-03
speaker        :: Taylor
outline        ::
Alignment comes up in several contexts and scenarios in MRI analyses
    + the purpose of the alignment (motion correct, going to standard space?) 
      determines choices in performing alignment
    + knowing the properties of datasets being aligned (contrasts, resolution) 
      matters a lot, as well
Note that afni_proc.py will help take care of *lots* alignment considerations
Important background concepts/vocabulary:
    + tissue contrasts    : which tissues are brightest/less bright?
    + spatial resolution  : how many voxels per area or volume?
    + field of view (FOV) : whole or partial brain coverage?
    + dataset "overlap"   : large rotation/translation/distortion between dsets?
Alignment requires quantifying similarity of dsets ("cost function")
Having a good+appropriate cost function matters a lot
    + knowing relative tissue contrasts determines cost function choice
      - AFNI's lpa* : dsets with similar contrasts
      - AFNI's lpc* : dsets with differing contrasts
      - (also, for 3dvolreg: default 'ls' for quick/good alignment across EPI)
    + different software use different cost functions    
Know how many degrees of freedom (DFs) to allow for a given type of alignment
    + rigid body    : used for motion "correction" (EPI-EPI)
    + linear affine : used for EPI-anatomical
    + nonlinear     : used for between-subject, or some distortion correction
Need to evaluate alignment visually: check how interior structures/gyri match
Alignment produces 'transformations' ('maps'), which can be applied to other dsets
    + e.g., calculate transform between anatomicals, and apply it to an ROI dset
Applying transformations always produces smoothing/interpolating
    + datasets are discrete grids, and transformations re-grid dsets
    + can choose appropriate kernel shape for final interpolation
    + e.g., wsinc5 preserves edges well; NN preserves integer values (ROIs)
When applying several transforms (i.e., EPI through motion corr, alignment
    to anatomical, and then to standard space), should *concatenate* transforms
    to apply only 1, rather than applying separately and regridding multiple times
    + creates less smoothing overall
    + afni_proc.py will take care of this for you!