help
help > RE: Problem comparing Results in SPM vs. Conn
Jul 1, 2014 10:07 PM | Alfonso Nieto-Castanon - Boston University
RE: Problem comparing Results in SPM vs. Conn
Hi Julia,
Thanks for reporting this. Both of these issues (the difference in reported peak voxel coordinates, and the problem when using *very* small voxel-level threshold p-values) arise from a limited-precision issue when computing p-values for very large T- or F- statistics (all voxels with T-values above a given threshold -this depends on the degrees of freedom- which roughly corresponds to p-values below 10^-15 are being treated as equally significant). I will fix this (by reporting peak voxels within each cluster using the highest T- or F- values instead of the lowest p-values) and add the ability to define a voxel-level threshold based on T- or F- stats in the next release of the toolbox (which should be out in a few days at most).
Regarding the missing clusters in the surface-projection display used in CONN (compared to SPM's surface-projection display), if you are referring to the cerebellar sources CONN surface-projection display only projects to the cortical surface (no subcortical structures since for these areas it is not particularly informative to project the results to their external surfaces). If you wish to display those clusters I would suggest using the 'volume-display' instead. We are working on finding better ways to combine surface-level results (for cortical results) with volume-level results (for subcortical structures), but in the meantime I will consider adding a subcortical structures 3d-render model to the volume-display just to have a more informative reference when displaying cerebellar or subcortical activations. In addition to this, there are several differences in the way these surface-projection displays are generated between CONN and SPM, the most obvious being the different cortical-surface mesh used both for extraction and display of the results (CONN uses freesurfer-generated fsaverage pial/white-matter surfaces for projecting your volume-based results to the surface; this should be more similar -but still not identical- to the new canonical cortical surfaces included in SPM8 rather than to those available in SPM5, which I believe are the ones you might be using in your attached figures). In any way, please keep in mind that, compared to performing surface-level analyses (which you can of course also do in CONN if you have run your subject structural volumes through freesurfer), projecting volume-based results on the cortical surface is only a convenient way to display the results, but it is never going be an extremely accurate representation because of the unavoidable uncertainties when matching MNI-normalized volume-based results to cortical-surface locations.
Hope this helps
Alfonso
Originally posted by Julia Landsiedel:
Thanks for reporting this. Both of these issues (the difference in reported peak voxel coordinates, and the problem when using *very* small voxel-level threshold p-values) arise from a limited-precision issue when computing p-values for very large T- or F- statistics (all voxels with T-values above a given threshold -this depends on the degrees of freedom- which roughly corresponds to p-values below 10^-15 are being treated as equally significant). I will fix this (by reporting peak voxels within each cluster using the highest T- or F- values instead of the lowest p-values) and add the ability to define a voxel-level threshold based on T- or F- stats in the next release of the toolbox (which should be out in a few days at most).
Regarding the missing clusters in the surface-projection display used in CONN (compared to SPM's surface-projection display), if you are referring to the cerebellar sources CONN surface-projection display only projects to the cortical surface (no subcortical structures since for these areas it is not particularly informative to project the results to their external surfaces). If you wish to display those clusters I would suggest using the 'volume-display' instead. We are working on finding better ways to combine surface-level results (for cortical results) with volume-level results (for subcortical structures), but in the meantime I will consider adding a subcortical structures 3d-render model to the volume-display just to have a more informative reference when displaying cerebellar or subcortical activations. In addition to this, there are several differences in the way these surface-projection displays are generated between CONN and SPM, the most obvious being the different cortical-surface mesh used both for extraction and display of the results (CONN uses freesurfer-generated fsaverage pial/white-matter surfaces for projecting your volume-based results to the surface; this should be more similar -but still not identical- to the new canonical cortical surfaces included in SPM8 rather than to those available in SPM5, which I believe are the ones you might be using in your attached figures). In any way, please keep in mind that, compared to performing surface-level analyses (which you can of course also do in CONN if you have run your subject structural volumes through freesurfer), projecting volume-based results on the cortical surface is only a convenient way to display the results, but it is never going be an extremely accurate representation because of the unavoidable uncertainties when matching MNI-normalized volume-based results to cortical-surface locations.
Hope this helps
Alfonso
Originally posted by Julia Landsiedel:
Dear all,
I encountered the following issue using the conn and SPM display. I used the exact same p-values and extent thresholds but surprisingly SPM outputs different MNI coordinates for the peak voxels (in a seed to voxel analysis) compared to conn. The coordinates are mostly just slightly shifted. How could that be? Did anybody encounter the same problem? Related to this comparing the surface display in conn and the render in SPM there are some activations not displayed in the conn display. Any idea for this? (see attached some screenshots for both problems)
Also I was wondering whether it could be possible to add to the toolbox functions that one also can be able to enter exact T-Values in the conn-display as height threshold. I also realized that using p-values so small that they extend a T-value of 10 makes all connectivity disappear in the conn display whereas in SPM it works fine. Also when going back to a higher p-value in the conn display afterwards gives the following warning:
Warning: Returning NaN for out of range arguments
> In spm_Pcdf at 90
In spm_P_RF at 110
In spm_P at 47
In conn_vproject>vproject_display at 737
In conn_vproject at 413
Thanks a lot!
Best,
Julia
I encountered the following issue using the conn and SPM display. I used the exact same p-values and extent thresholds but surprisingly SPM outputs different MNI coordinates for the peak voxels (in a seed to voxel analysis) compared to conn. The coordinates are mostly just slightly shifted. How could that be? Did anybody encounter the same problem? Related to this comparing the surface display in conn and the render in SPM there are some activations not displayed in the conn display. Any idea for this? (see attached some screenshots for both problems)
Also I was wondering whether it could be possible to add to the toolbox functions that one also can be able to enter exact T-Values in the conn-display as height threshold. I also realized that using p-values so small that they extend a T-value of 10 makes all connectivity disappear in the conn display whereas in SPM it works fine. Also when going back to a higher p-value in the conn display afterwards gives the following warning:
Warning: Returning NaN for out of range arguments
> In spm_Pcdf at 90
In spm_P_RF at 110
In spm_P at 47
In conn_vproject>vproject_display at 737
In conn_vproject at 413
Thanks a lot!
Best,
Julia
Threaded View
| Title | Author | Date |
|---|---|---|
| Julia Landsiedel | Jul 1, 2014 | |
| Alfonso Nieto-Castanon | Jul 1, 2014 | |
| Alfonso Nieto-Castanon | Jul 2, 2014 | |
| Julia Landsiedel | Jul 3, 2014 | |