help
help > RE: Compcor:Is linear trend removed prior to pca?
Nov 16, 2015 06:11 PM | Alfonso Nieto-Castanon - Boston University
RE: Compcor:Is linear trend removed prior to pca?
Hi Immanuel,
Regarding (1) by default, when using CONN preprocessing pipeline, all ROI-extraction steps (including WM/CSF PCA decomposition) will be performed on the non-smoothed functional volumes. This is controlled by two elements in the GUI: a) in Setup.functional, the 'Functional data for ROI-level analyses' field will be set to 'Other: same files without leading 's' (SPM convention for unsmoothed volumes)'; and b) in Setup.ROIs, all of the ROIs will have the 'Use ROI source data' field checked. As long as those settings are kept, and the wau*.nii files generated as part of the preprocessing steps have not been deleted or moved to a different folder, when extracting BOLD signal timeseries from any ROI CONN will use the unsmoothed functional volumes.
Regarding (2), typically between 3 to 5 components per ROI seem to be "optimal" (this is derived from the analyses described in the Chai et al. 2012 Neuroimage paper). The criterion for selecting the number of components here is not really tied to the percentage explained variance within the noise ROIs, but rather on their impact on the sensitivity and specificity of the connectivity values after denoising. The general "rule of thumb" is to look at the resulting voxel-to-voxel correlation distribution and make sure that: a) the resulting distributions are reasonably centered and consistent across subjects (if they are not, then a more aggressive form of denoising is needed, e.g. adding more components to the CompCor ROIs, using a more conservative set of thresholds in ART, etc.); and b) the resulting degrees of freedom are not "too low" (if they are, then a less aggressive form of denoising is needed, e.g. reducing the number of components in the CmopCor ROIs, using a more liberal set of threshold in ART, etc.)
Hope this helps
Alfonso
Originally posted by Immanuel Elbau:
Regarding (1) by default, when using CONN preprocessing pipeline, all ROI-extraction steps (including WM/CSF PCA decomposition) will be performed on the non-smoothed functional volumes. This is controlled by two elements in the GUI: a) in Setup.functional, the 'Functional data for ROI-level analyses' field will be set to 'Other: same files without leading 's' (SPM convention for unsmoothed volumes)'; and b) in Setup.ROIs, all of the ROIs will have the 'Use ROI source data' field checked. As long as those settings are kept, and the wau*.nii files generated as part of the preprocessing steps have not been deleted or moved to a different folder, when extracting BOLD signal timeseries from any ROI CONN will use the unsmoothed functional volumes.
Regarding (2), typically between 3 to 5 components per ROI seem to be "optimal" (this is derived from the analyses described in the Chai et al. 2012 Neuroimage paper). The criterion for selecting the number of components here is not really tied to the percentage explained variance within the noise ROIs, but rather on their impact on the sensitivity and specificity of the connectivity values after denoising. The general "rule of thumb" is to look at the resulting voxel-to-voxel correlation distribution and make sure that: a) the resulting distributions are reasonably centered and consistent across subjects (if they are not, then a more aggressive form of denoising is needed, e.g. adding more components to the CompCor ROIs, using a more conservative set of thresholds in ART, etc.); and b) the resulting degrees of freedom are not "too low" (if they are, then a less aggressive form of denoising is needed, e.g. reducing the number of components in the CmopCor ROIs, using a more liberal set of threshold in ART, etc.)
Hope this helps
Alfonso
Originally posted by Immanuel Elbau:
Hi Alfonso,
I have two further questions concerning the aCompCor implementation in CONN:
1) As a default, when one uses the CONN preprocessing pipeline, is the pca performed on spatially smoothed images or on non-smoothed ones?
2) For CSF and WM how-many components Do you recommend to use as covariates? Is there a way to display the amount of cumulative variance explained for each component within CONN to inform this decision?
Thank you and all the best!
Immanuel
I have two further questions concerning the aCompCor implementation in CONN:
1) As a default, when one uses the CONN preprocessing pipeline, is the pca performed on spatially smoothed images or on non-smoothed ones?
2) For CSF and WM how-many components Do you recommend to use as covariates? Is there a way to display the amount of cumulative variance explained for each component within CONN to inform this decision?
Thank you and all the best!
Immanuel
Threaded View
| Title | Author | Date |
|---|---|---|
| Immanuel Elbau | Sep 17, 2015 | |
| Alfonso Nieto-Castanon | Sep 17, 2015 | |
| Immanuel Elbau | Nov 15, 2015 | |
| Alfonso Nieto-Castanon | Nov 16, 2015 | |
| Immanuel Elbau | Oct 21, 2015 | |