Hi there,
I am currently using the FATCAT commands in order to combine fmri and dti data.
However, I noticed that after performing the dual regression (on the functional data), in order to get z-score maps at the subject level, each subject contains a different number of ROIs (3dROIMaker step). We don’t get the same number of ROI’s across all subjects. Do you have any suggestions on how I can overcome this?
I was wondering if there was some way around this issue. I know with 3dNetCorr there is a “push_thru_zeros” option but how about the DTI side? Is there a way to run 3dtrackID without getting errors based on the different number of ROI’s?
Also, how should I deal with a different number of ROI’s per subject, when it comes to statistical analysis?
Thanks for your time.
Hi, Sondos-
The heterogeneity of such (esp. since the individual level results are usually quite noisy in FMRI) can make challenges. Note that having different numbers of ROIs won’t cause errors in 3dNetCorr or 3dTrackID per se; however, trying to combine matrices of different sizes after those steps would present challenges, indeed. Those analyses are typically based on looking at “connections” between the same two targets across a group (whether those be correlation values or tract-derived estimates); therefore, having different targets to start with would not really fit with this approach.
Two ways to go might be:
- from the individual-level results, only keep ROIs that are common across subjects.
- map the group-level ROIs to each subject and use those for tracking/netcorring.
–pt
Thank you for your response.
I see what you’re saying but I had questions with both approaches you suggested.
for the first suggestion: 1) from the individual-level results, only keep ROIs that are common across subjects.
I was trying to do that, but I was saying how about in the case where some individual-level results contain fewer number of ROI’s compared to the group? Even if I try to reduce the number of ROI’s for each subject so that only the common ROI’s are left. What could end up happening is that one of the subjects has only one ROI… so in that case do you suggest I reduce the number of ROIs to match that? Even though the group-level ROI’s have 5 ROI’s?
and for the second suggestion: 2) map the group-level ROIs to each subject and use those for tracking/netcorring.
In this case wouldn’t we be losing some activation information in the process? Every brain is shaped differently and activates differently, so if we were to overlay the group results onto the individual subject level, we may be losing valuable information.
So what I’ve been trying to do is 1. apply melodic (GICA) and get from that the independent components of the entire group of subjects. I then perform 3dMatch on the group level. After that I perform 3dROIMaker for the group in order to get different number of ROI’s for each network.
Once I have the ROI’s for each network at the group level, I perform dual regression using my 3dROIMaker output and the individual subjects. I then peform 3dROIMaker once again, but this time to convert the individual-level z-score maps obtained from dual regression into ROI’s at the individual level. Here is where my problem comes in and the number of ROI’s vary across subjects.
So from this step, that all the DMN networks for each individual level subject contains 5 ROI’s for instance. And if they contain over 5 ROI’s, I remove the extra ROI’s. But then what should I do if there are fewer than 5 ROI’s for each subject? let’s say some subjects have 4, 3 or even 2 ROI’s?
Thanks for your help
Hi, Sondos-
Re.:
for the first suggestion: 1) from the individual-level results, only keep ROIs that are common across subjects.
I was trying to do that, but I was saying how about in the case where some individual-level results contain fewer number of ROI’s compared to the group? Even if I try to reduce the number of ROI’s for each subject so that only the common ROI’s are left. What could end up happening is that one of the subjects has only one ROI… so in that case do you suggest I reduce the number of ROIs to match that? Even though the group-level ROI’s have 5 ROI’s?
yep, that just seems like an inherent difficult-- FMRI data is noisy, and so individual subject results can vary a lot. Using the individual subject ROIs can lead to such difficulties.
Re.:
and for the second suggestion: 2) map the group-level ROIs to each subject and use those for tracking/netcorring.
In this case wouldn’t we be losing some activation information in the process? Every brain is shaped differently and activates differently, so if we were to overlay the group results onto the individual subject level, we may be losing valuable information.
Well, isn’t that an inherent difficulty with alignment? ICA is finding group average results anyways. Any method of combining subjects for a group analysis will lose some individual information in the “group test” part, won’t it?
Re.:
So what I’ve been trying to do is 1. apply melodic (GICA) and get from that the independent components of the entire group of subjects. I then perform 3dMatch on the group level. After that I perform 3dROIMaker for the group in order to get different number of ROI’s for each network.
Makes sense
Re.:
Once I have the ROI’s for each network at the group level, I perform dual regression using my 3dROIMaker output and the individual subjects. I then peform 3dROIMaker once again, but this time to convert the individual-level z-score maps obtained from dual regression into ROI’s at the individual level. Here is where my problem comes in and the number of ROI’s vary across subjects.
I don’t understand this part. Didn’t you get your ROIs from the group ICA stage just above?
Re.:
So from this step, that all the DMN networks for each individual level subject contains 5 ROI’s for instance. And if they contain over 5 ROI’s, I remove the extra ROI’s. But then what should I do if there are fewer than 5 ROI’s for each subject? let’s say some subjects have 4, 3 or even 2 ROI’s?
Going back to the initial difficulties with the first method mentioned above, I think using individual-derived ROIs in this way might be tricky. ICA results on noisy individual subject data will likely vary a lot.
–pt