title          :: Alignment (part 2/4): EPI-EPI
presentation   :: afni14_alignment.pdf
scripts        :: afni14_alignment_cmds_3dvolreg.txt
date           :: 2020-03
speaker        :: Taylor
outline        :: 
EPI-EPI alignment, mainly to address motion within subject time series
Within-volume motion is can be easily seen in sagittal slices
    + bright/dark pattern appears, due to interleaved acquisition
Motion causes spikes in time series, sometimes up or down
AFNI's 3dvolreg can be used to estimate motion/alignment parameters across time
    + rigid body registration (default cost function: 'ls')
    + get a time series of motion parameters, relative to one reference 'base' vol
    + time series can be used as regressors in GLM
      - not perfect to eliminate motion because spikes go up/down unpredictably
    + can censor bad volumes based on motion estimates
      - 'enorm' (Euclidean norm) combines the motion parameters into single value
        - where enorm is large, motion event was large
        - combines *differences* of parameters across time
        - if enorm is greater than pre-defined threshold (e.g., 0.2 mm), censor
        - enorm censoring removes *2* volumes (bc from "differences" of parameters)
      - outlier fractions show where a volume had lots of spikes/outliers in time
        - calculate outliers before any motion correction or processing
        - if volume has too large fraction of outliers (e.g., GT 5%), can censor it
      - afni_proc.py QC shows these values and censoring results
(Hands-on example using 3dvolreg, looking at motion correction and enorm)
There are lots of caveats with motion correction
    + can never make it look like no motion occurred in the data!
    + always check output of 3dvolreg to verify!
    + design your study carefully---don't want task-correlated motion
    + be careful motion differences doesn't affect group results
(There are EPI-distortion correction tools in AFNI/afni_proc.py
    + nonlinear alignment of blip up/down EPIs
    + B0/phase dsets can be applied)
Global signal regression is *not* recommended to deal with motion
    + see the large number of papers on this...
    + 'GCOR' is a better way to go at group level
Plan/practice your study design carefully!
    + MUCH better to not introduce motion than to try to fix later.