Is it possible to save the correlation matrix from the voxel to voxel analysis in a "standard format" for additional analysis with other software packages?

From the Conn Toolbox paper "the correlation matrix is stored as matrix could occupy above 300 Gb for each subject), this matrix is instead characterized without loss of precision by its eigenvectors and associated eigenvalues (see Appendix A.2)."   I am hoping that I will be able to reduce the memory requirements by increasing the voxel size to something reasonable.   I understand how the image should be reconstructed from the eigenvalues and eigenvectors but not exactly sure on the implementation.

I am hoping that 

1) code already exists to do this conversion in the CONN Toolbox and/or 
2) More information as to where and how the individual eigenvalues and eigenvectors are stored.

Thanks in advance for everyone's help and comments.

Bob

I came here to ask the exact same question.

I do know a first-level analysis produces a matlab file in the results/firstlevel folder, and that this matlab data file contains a matrix called C (64x64). Could this matrix perhaps somehow be used to reconstruct the correlation matrix?

Thanks in advance

Hi Bob & Michael,

Unfortunately currently there is no user-level code to do this conversion in the CONN toolbox (from the internal SVD-based representation of the voxel-to-voxel correlations used in "voxel-to-voxel" analyses, to the actual voxel-to-voxel correlation matrix). The reason is that this internal representation is used precisely to avoid having to do that conversion (which is very demanding both computationally and in terms of required storage), so all "voxel-to-voxel" analyses are actually transformed and performed in this internal representation in order to improve performance.

What I would suggest is to use a relatively coarse-grained parcellation of the brain and perform ROI-to-ROI analyses instead. That will give you readily the ROI-to-ROI connectivity matrices that you want. For example, if you were willing to use 5mm "voxels" to reduce the dimensionality of your voxel-to-voxel matrix, simply create an ROI file that defines 5mm cube ROIs covering the entire brain, and then perform ROI-to-ROI analyses using these custom-made ROIs. That will also have the advantage of allowing you then to perform all sort of connectome level analyses on your results within the CONN toolbox (e.g. graph theoretic measures, network-based statistics, etc.)

Hope this helps
Alfonso


Originally posted by Bob Kraft:

Alfonso,

Thank you very much for this suggestion. I will give it a try and let you know how it goes.

Bob