External Email - Use Caution
Dear FreeSurfer experts,
In FreeSurfer (Darwin-OSX-stable-v6-beta-20151015 build, operating on Yosemite 10.10.5) I’m running into an issue that I would like to request your expertise on.
The procedure I used:
I have a dataset of 77 participants, which have been processed through the standard recon-all and qcache pipeline. Subsequently I’ve used mris_preproc to prepare the data for group anaylysis, using FSGD files with the DeMeanFlag 1 and ReScaleFlag 1 options. After that, MRI_glmfit was run with the appropriate contrasts. Finally, mri_glmfit-sim was used to apply cluster correction. The only (minor?) issue until this point is that mri_glmfit-sim often runs into a “Segmentation fault” error, which can be bypassed by simply rerunning mri_glmfit-sim until it does work. Having done this, there seem to be no further issues and the clusters resulting from the glm analysis, also after cluster correction, look nice in both hemispheres.
To further interpret the resulting clusters, I extracted individual labels from the annotation files resulting from the glm analyses, using mri_annotation2label. These labels were checked and looked correct. Next, these fsaverage-based labels were then warped back to the individual level, using mri_label2label, with each subject as the –trgsubject. The registration between individual labels and their respective inflated brain was checked and looked good on every hemisphere. At this point mris_anatomical_stats was used to extract average cortical thickness data for each cluster in each participant. Finally, using aparcstats2table, these values were placed in a table to be imported in R statistics software for visualization purposes.
What the main problem is:
The main issue I would like to ask your advice on appeared when the data was analyzed in R. In R, the same regression models were reconstructed as the ones that were used in mri-glmfit, in order to parse and visualize the contribution of each predictor in the model to the cortical thickness measure in each cluster. In the right hemisphere, the results were as expected for every cluster, showing highly significant effects of the variables of interest (p<0.001). However, when the same was repeated for the clusters in the left hemisphere, none of the expected effects were seen, with p values easily rising above 0.90. As the clusters in the left hemisphere survived mri_glmfit-sim cluster correction, and the outcome measure is average cortical thickness extracted from these clusters, there should be no way that this is correct data.
The fact that the data looks incorrect specifically in the left hemisphere, while they are as expected for the right hemisphere, leads me to believe left and right hemisphere may have been mixed up in one of the scripts. However, I have checked the used scripts meticulously for any point at which left and right hemisphere may have accidentally been mixed up, but after repeated checks I have not found any issues. The procedures from mri_label2label and onwards were also repeated with separate scripts for left and right hemispheres as a sanity check, and the resulting thickness values did not change. As mentioned before, the label overlays look good for both hemispheres at the individual level, so I do not believe anything went wrong at this point.
I would like to ask if you have any suggestions as to where you suspect the problem might lie and would greatly appreciate any pointers you might have.
If any more information is needed, please let me know.
Kindest regards,
Roy
You should try the analysis but with the data in the something.y.ocn.dat output by mri_glmfit-sim. This data contains a nsubjects x nclusters table of the input data (y) to the glm. If that does not show up as significant, then something is wrong. doug
On 2/5/2020 11:30 AM, Hooren, Roy van (Alumni) wrote:
External Email - Use Caution
Dear FreeSurfer experts,
In FreeSurfer (Darwin-OSX-stable-v6-beta-20151015 build, operating on Yosemite 10.10.5) I’m running into an issue that I would like to request your expertise on.
*The procedure I used:*
I have a dataset of 77 participants, which have been processed through the standard recon-all and qcache pipeline. Subsequently I’ve used mris_preproc to prepare the data for group anaylysis, using FSGD files with the DeMeanFlag 1 and ReScaleFlag 1 options. After that, MRI_glmfit was run with the appropriate contrasts. Finally, mri_glmfit-sim was used to apply cluster correction. The only (minor?) issue until this point is that mri_glmfit-sim often runs into a “Segmentation fault” error, which can be bypassed by simply rerunning mri_glmfit-sim until it does work. Having done this, there seem to be no further issues and the clusters resulting from the glm analysis, also after cluster correction, look nice in both hemispheres.
To further interpret the resulting clusters, I extracted individual labels from the annotation files resulting from the glm analyses, using mri_annotation2label. These labels were checked and looked correct. Next, these fsaverage-based labels were then warped back to the individual level, using mri_label2label, with each subject as the –trgsubject. The registration between individual labels and their respective inflated brain was checked and looked good on every hemisphere. At this point mris_anatomical_stats was used to extract average cortical thickness data for each cluster in each participant. Finally, using aparcstats2table, these values were placed in a table to be imported in R statistics software for visualization purposes.
*What the main problem is:*
The main issue I would like to ask your advice on appeared when the data was analyzed in R. In R, the same regression models were reconstructed as the ones that were used in mri-glmfit, in order to parse and visualize the contribution of each predictor in the model to the cortical thickness measure in each cluster. In the right hemisphere, the results were as expected for every cluster, showing highly significant effects of the variables of interest (p<0.001). However, when the same was repeated for the clusters in the left hemisphere, none of the expected effects were seen, with p values easily rising above 0.90. As the clusters in the left hemisphere survived mri_glmfit-sim cluster correction, and the outcome measure is average cortical thickness extracted from these clusters, there should be no way that this is correct data.
The fact that the data looks incorrect specifically in the left hemisphere, while they are as expected for the right hemisphere, leads me to believe left and right hemisphere may have been mixed up in one of the scripts. However, I have checked the used scripts meticulously for any point at which left and right hemisphere may have accidentally been mixed up, but after repeated checks I have not found any issues. The procedures from mri_label2label and onwards were also repeated with separate scripts for left and right hemispheres as a sanity check, and the resulting thickness values did not change. As mentioned before, the label overlays look good for both hemispheres at the individual level, so I do not believe anything went wrong at this point.
I would like to ask if you have any suggestions as to where you suspect the problem might lie and would greatly appreciate any pointers you might have.
If any more information is needed, please let me know.
Kindest regards,
Roy
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Douglas N. Greve -
Hooren, Roy van (Alumni)