01 July 2009
AFNI Realtime - for AJ and JH at MCW
1) Input multi-coil complex images = Cx# (where #=0,1,2,...)
2) form absolute sum volume = AbsSum
3) register AbsSum in time = AbSumReg
4) apply registration parameters to each Cx# volume = CxReg#
Before the above, maybe also
A) Input B0 map
B) Smooth, convert to EPI displacement field
C) Unwarp Cx# images in y-direction before further processing
Main drawbacks -- this is really boring to do
2) form absolute sum volume = AbsSum
3) register AbsSum in time = AbSumReg
4) apply registration parameters to each Cx# volume = CxReg#
Before the above, maybe also
A) Input B0 map
B) Smooth, convert to EPI displacement field
C) Unwarp Cx# images in y-direction before further processing
Main drawbacks -- this is really boring to do
AFNI Idea - Group Analysis
Related to the last post:
1) Create IM betas for each event/block
2) Compute means and SDs for these
3) These are the inputs to 3dMEMA instead of the t-statistics from 3dREMLfit
Advantage: allow for intra-subject performance variability
1) Create IM betas for each event/block
2) Compute means and SDs for these
3) These are the inputs to 3dMEMA instead of the t-statistics from 3dREMLfit
Advantage: allow for intra-subject performance variability
AFNI Idea - Connectivity in Task datasets
To get task-based connectivity on an ROI level:
1) Compute IM betas for each block/event
2) Test the collection of IM betas for nonzero mean (for each task class)
3) Threshold and cluster the voxels
4) Between each pair of clusters/ROIs, compute the correlation of IM betas in some way
Result is an activation map (from 3) and a 'functional connectivty' map between the active clusters. No dynamics, of course, or causality.
Significance of correlations? What's the null hypothesis here? Not just that there is no activation at all. Must be some activation + no-ROI-correlation hypothesis.
1) Compute IM betas for each block/event
2) Test the collection of IM betas for nonzero mean (for each task class)
3) Threshold and cluster the voxels
4) Between each pair of clusters/ROIs, compute the correlation of IM betas in some way
Result is an activation map (from 3) and a 'functional connectivty' map between the active clusters. No dynamics, of course, or causality.
Significance of correlations? What's the null hypothesis here? Not just that there is no activation at all. Must be some activation + no-ROI-correlation hypothesis.
AFNI Idea - Resting State
Resting state group analysis:
1) Compute correlation maps of all GM voxels with all other GM voxels = 10-30K volumes per subject
2) Convert r to z and then do a group t-statistic set of maps
3) Threshold on t and then cluster each map
4) Compute p-value for each surviving cluster
5) Do FDR q computations on all these p-values
Sub-ideas and problems:
1a) 'Smooth' with projection onto local SVD space of dimension x (small x)
1b) Remove first SVD vector of all non-GM voxels in a sphere about each GM voxel
1c) Use a sparse BCC mesh of seed voxels to cut down the number of maps
3a) What t threshold to use?
4a) Lots of overlapping clusters -- how to deal with these?
5a) With overlapping clusters (e.g., from nearby seed voxels), will have many similar small p-values that are highly correlated in fact. What do do? (same as 4a)
1) Compute correlation maps of all GM voxels with all other GM voxels = 10-30K volumes per subject
2) Convert r to z and then do a group t-statistic set of maps
3) Threshold on t and then cluster each map
4) Compute p-value for each surviving cluster
5) Do FDR q computations on all these p-values
Sub-ideas and problems:
1a) 'Smooth' with projection onto local SVD space of dimension x (small x)
1b) Remove first SVD vector of all non-GM voxels in a sphere about each GM voxel
1c) Use a sparse BCC mesh of seed voxels to cut down the number of maps
3a) What t threshold to use?
4a) Lots of overlapping clusters -- how to deal with these?
5a) With overlapping clusters (e.g., from nearby seed voxels), will have many similar small p-values that are highly correlated in fact. What do do? (same as 4a)
