[color=#000000]Hi Alfonso,[/color]
[color=#000000]RE: point number 2:[/color]
If there are, for example, 14 voxels at 1mm res, will CONN force to obtain signal from 14 voxels at 2mm even if those coordinates are not in that region? Or would it pull from fewer voxels so that the anatomical specificity remains intact?
Thank you,
Sarah
[i]Originally posted by Alfonso Nieto-Castanon:[/i][quote][color=#000000]Hi Ely,[/color]
[color=#000000]Those are very good questions, some thoughts on these below[/color]
[color=#000000]Best[/color]
[color=#000000]Alfonso[/color]
[i]Originally posted by Benjamin Ely:[/i][quote]Hi Alfonso,
I'm working on a CONN analysis (version 15a) that uses several small (2-5 functional voxels), subject-specific ROIs. I have a couple of questions:
1) Is it an issue that the subject-specific ROIs I've generated are already in MNI space? I do most of my preprocessing (i.e. realignment, warping to MNI, smoothing) outside of CONN, so all the structural/functional/ROI files I uploaded into the setup page were already in MNI space. I just noticed from the manual that subject-specific ROIs should be in subject space. Is CONN applying an additional MNI transformation to the ROIs I uploaded, and if so, can I disable this?
[/quote]
That is perfectly fine, and sorry if the manual is not perfectly clear on this regard. CONN does not apply any transformation at all to your ROIs, so the ROI files that you enter into CONN in Setup.ROIs (whether subject-specific or subjet-independent) are expected to always be coregistered (in the same space, but of course not necessarily resampled to the same resolution) as the corresponding functional volumes that you enter in Setup.functionals (after any preprocessing steps if applicable). So, basically, after any preprocessing step and right before running the Setup step CONN expects all of the files (functionals, structurals, ROIs) to be already appropriately coregistered. The manual remark concerning subject-specific ROIs is just referring to the case (not yours) where you have subject-specific ROIs [i]in subject-space [/i](e.g. defined anatomically from the original -before normalization- structural volumes), and of course in that case you should make sure that your associated functional files are also in the same space. [quote]
2) How does CONN interpolate higher-resolution ROIs to functional resolution? The ROIs I have were created in anatomical space (0.7mm iso); I did my own downsampling to functional space (2mm iso) to ensure fidelity and imported both sets during setup. The results between the two look quite similar, but not identical.
[/quote]
In this regard CONN will always respect the original ROI files resolution. When extracting functional data from an ROI CONN will first get the coordinates of all of the voxels within the ROI (at the resolution of the ROI file), and then extract the functional data at these same coordinates from the functional volumes using nearest neighbor interpolation. So, for example, if an ROI contains 5 voxels (at 0.7mm resolution), CONN will get the MNI coordinates of those 5 voxels and extract the BOLD timeseries from the 5 voxels in the functional data that are closest to these 5 coordinate positions (which may all be from the same voxel in the functional data or from several voxels). After getting these 5 timeseries CONN will compute the average (or PCA for multidimensional ROIs like White/CSF areas) across these 5 timeseries to get the ROI-level BOLD timeseries. The reason it is done this way (instead of the opposite way: sampling the ROI files at the functional data resolution) is because the latter approach can lead to loss of very small ROIs due to resampling, while the former approach guarantees no loss of small ROIs and a more appropriate partial-volume weighting of the functional data. [quote]
3) In a similar vein, how does CONN decide which functional voxels to exclude when using the grey matter masking option? The grey matter mask I input (generated from the standard SPM segmentation program) is graded, not binary, and is also at anatomical resolution. I'm a bit unclear on how CONN translates this information into a binary grey matter mask at functional resolution.
[/quote]
Same principle as above. The grey matter mask values are sampled at the coordinates of the ROI-file voxels (in the example above we get 5 values from the grey matter files extracted from the voxels in those files that are closest to the coordinates of the 5 ROI voxels), and any ROI voxels outside of the mask are disregarded. CONN uses a relatively conservative threshold/masking for the 'mask with grey matter voxels' ROI option (only removes voxels with 0 values in the gray-matter mask,, despite the gray-matter mask generated by SPM segmentation being graded). This is mainly because the transients between 1-values and 0-values in these volumes are typically relatively fast (e.g. removing only the 0-values results in perhaps only ~20% more voxels within the "grey matter" mask than removing values below .5) and because for this masking (compared to those masks used for White-matter and CSF areas to be used in CompCor, for example, that benefit from a more aggresive approach -for those CONN uses instead a <.5 threshold followed by an additional erosion step) a more conservative approach is probably preferable (but of course if you prefer a more agressive masking you may do so simply by thresholding the corresponding c1*.nii files using any desired approach before running the Setup step).
Hope this helps
Alfonso[/quote]
[color=#000000]RE: point number 2:[/color]
If there are, for example, 14 voxels at 1mm res, will CONN force to obtain signal from 14 voxels at 2mm even if those coordinates are not in that region? Or would it pull from fewer voxels so that the anatomical specificity remains intact?
Thank you,
Sarah
[i]Originally posted by Alfonso Nieto-Castanon:[/i][quote][color=#000000]Hi Ely,[/color]
[color=#000000]Those are very good questions, some thoughts on these below[/color]
[color=#000000]Best[/color]
[color=#000000]Alfonso[/color]
[i]Originally posted by Benjamin Ely:[/i][quote]Hi Alfonso,
I'm working on a CONN analysis (version 15a) that uses several small (2-5 functional voxels), subject-specific ROIs. I have a couple of questions:
1) Is it an issue that the subject-specific ROIs I've generated are already in MNI space? I do most of my preprocessing (i.e. realignment, warping to MNI, smoothing) outside of CONN, so all the structural/functional/ROI files I uploaded into the setup page were already in MNI space. I just noticed from the manual that subject-specific ROIs should be in subject space. Is CONN applying an additional MNI transformation to the ROIs I uploaded, and if so, can I disable this?
[/quote]
That is perfectly fine, and sorry if the manual is not perfectly clear on this regard. CONN does not apply any transformation at all to your ROIs, so the ROI files that you enter into CONN in Setup.ROIs (whether subject-specific or subjet-independent) are expected to always be coregistered (in the same space, but of course not necessarily resampled to the same resolution) as the corresponding functional volumes that you enter in Setup.functionals (after any preprocessing steps if applicable). So, basically, after any preprocessing step and right before running the Setup step CONN expects all of the files (functionals, structurals, ROIs) to be already appropriately coregistered. The manual remark concerning subject-specific ROIs is just referring to the case (not yours) where you have subject-specific ROIs [i]in subject-space [/i](e.g. defined anatomically from the original -before normalization- structural volumes), and of course in that case you should make sure that your associated functional files are also in the same space. [quote]
2) How does CONN interpolate higher-resolution ROIs to functional resolution? The ROIs I have were created in anatomical space (0.7mm iso); I did my own downsampling to functional space (2mm iso) to ensure fidelity and imported both sets during setup. The results between the two look quite similar, but not identical.
[/quote]
In this regard CONN will always respect the original ROI files resolution. When extracting functional data from an ROI CONN will first get the coordinates of all of the voxels within the ROI (at the resolution of the ROI file), and then extract the functional data at these same coordinates from the functional volumes using nearest neighbor interpolation. So, for example, if an ROI contains 5 voxels (at 0.7mm resolution), CONN will get the MNI coordinates of those 5 voxels and extract the BOLD timeseries from the 5 voxels in the functional data that are closest to these 5 coordinate positions (which may all be from the same voxel in the functional data or from several voxels). After getting these 5 timeseries CONN will compute the average (or PCA for multidimensional ROIs like White/CSF areas) across these 5 timeseries to get the ROI-level BOLD timeseries. The reason it is done this way (instead of the opposite way: sampling the ROI files at the functional data resolution) is because the latter approach can lead to loss of very small ROIs due to resampling, while the former approach guarantees no loss of small ROIs and a more appropriate partial-volume weighting of the functional data. [quote]
3) In a similar vein, how does CONN decide which functional voxels to exclude when using the grey matter masking option? The grey matter mask I input (generated from the standard SPM segmentation program) is graded, not binary, and is also at anatomical resolution. I'm a bit unclear on how CONN translates this information into a binary grey matter mask at functional resolution.
[/quote]
Same principle as above. The grey matter mask values are sampled at the coordinates of the ROI-file voxels (in the example above we get 5 values from the grey matter files extracted from the voxels in those files that are closest to the coordinates of the 5 ROI voxels), and any ROI voxels outside of the mask are disregarded. CONN uses a relatively conservative threshold/masking for the 'mask with grey matter voxels' ROI option (only removes voxels with 0 values in the gray-matter mask,, despite the gray-matter mask generated by SPM segmentation being graded). This is mainly because the transients between 1-values and 0-values in these volumes are typically relatively fast (e.g. removing only the 0-values results in perhaps only ~20% more voxels within the "grey matter" mask than removing values below .5) and because for this masking (compared to those masks used for White-matter and CSF areas to be used in CompCor, for example, that benefit from a more aggresive approach -for those CONN uses instead a <.5 threshold followed by an additional erosion step) a more conservative approach is probably preferable (but of course if you prefer a more agressive masking you may do so simply by thresholding the corresponding c1*.nii files using any desired approach before running the Setup step).
Hope this helps
Alfonso[/quote]