I ran a pre-processing pipeline that gives, for task data, stats files for data processed both with and without ANATICOR (-regress_anaticor_fast).
From my understanding ANATICOR uses WM masks to estimate noise coming from e.g. the scanner and coils. It also should reduce white matter clusters (right?).
When running a group analysis using 3dMVM on 36 subjects and looking at an interaction of interest I find that the “non_anaticor” data looks a bit cleaner and less clusters in white matter. Attaching a few prints (top row: anaticor, bottom row: no anaticor). These subjects are alcoholics which means that some of them have less WM and some atrophy. Could this affect the performance of anaticor and discourage us from use it in subjects with atrophy?
As you can see the GM pre-frontal cluster is bigger and stronger without anaticor (bottom row) and the WM clusters (in the cross hair) is only there for anaticor (top row). The insula cluster is only visible in the anaticor data but it is close to WM and you have to lower the threhold a lot to show that cluster in the non_anaticor data. Why would the results be so profoundly different?
So: The results differ a lot between anaticor and none anaticor plus that it is not apparent that anaticor provides the best results. In a group map it even gives more WM clusters at the same threshold.
You ask good questions. To which I don’t really have answers. We’ve never dealt with using ANATicor on people with WM atrophy. In such a case, I would definitely not rely on the built-in AFNI 3dSeg program for finding the WM – at least without some looking at its results to make sure they are “reasonable” – and how one defines “reasonable” in this context I don’t know.
I can’t say we’ve seen the creation of WM blobs by ANATicor. If this is common, but only in people with WM atrophy, then that is also an interesting effect. Does it also occur if you use the COMPcor-like analysis method available in afni_proc.py instead of ANATicor? (That is the “-regress_ROI_PC” option – I’ve never used this option myself, so if you want to try it, you’ll have to pester Rick Reynolds.)
I did some further digging. I created an average map of the “mask_WMe_resam +tlrc” file for ca 30 subjects from both the group we overall expect to have more atrophy issues and one group of healthy adolescents. We then compared them to each other and to the WMe_resam that we get from running 3dSeg + 3dmask_tool + 3dresample on the TT_N27+tlrc template.
What we can say is this:
The WMe_resam that you get from the TT_N27+tlrc set is pretty good. The loss in Corpus Callosum is due to the re-sampling.
Overall (both healthy adolescents and alcoholics with expected atrophy) labels some sub-cortical structures as WM, e.g. Putamen (see the images below).
Corpus Callosum is an example where the healthy adolescents overall get a more correct WM labelling and the patients gets an extended WM segmentation where GM dorsally of Corpus Callosum gets labelled at WM. Look at the diff_map where red means patients have a higher average white matter value and blue means that the value was higher in the healthy group average WM map.
By overlaying the diff_map over the WM segmented out of the TT_N27+tlrc tempalte we can see that the blue “healthy” voxels better overlap with the white TT_N27_WMe_resam underlay and that the red “patient” voxels more often extend the WM structures.
Difficult followup question: How could these findings explain the results I posted previously? Do you recommend not running anaticor on these patients? Since the WM mask covers all of Putamen in many subjects we might loose interesting signal from there? Also true for the healthy group though but not as bad.
Images: attached as a google drive link since there is a limit of 2 files to uppload and you cant upload .zip files:
TT_seg_resam = Segmentation preformed on the TT_N27+tlrc template where I resampled to the resolution of the EPIs in our experiment. The segmentation seems to be good.
healthy_mean = 3dmean run on the mask_WMe_resam of 30 healthy subjects (values 0-1) overlayed on TT_N27+tlrc. Looks decent but it covers a lot of Putamen and some GM areas.
patients_mean = 3dmean run on the mask_WMe_resam of 30 alcoholic patients subjects (values 0-1) overlayed on TT_N27+tlrc. It seems like it tries to categorize more GM as WM and Putamen is more often classified as WM.
mean_diff = mean_patients - mean_healthy. Red means that the region is overall more commonly classifed as WM in the patient group and blue means that the region is more commonly classifed as WM in the healthy group. Putamen and the GM dorsally of Corpus Callosum is clearly more often labeled as WM in the patient group.
TT_WMe_resam_diff = The difference map overlayed over the the WMe_resam of the TT_N27 with lowered opacity. Here it is quite clear that the blue “healthy” voxels better overlap with the white TT_N27_WMe_resam underlay and that the red “patient” voxels more commonly expand over the TT_N27_WMe_resam underlay, indicating that more GM is clasified as WM in the patients.
The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.
