iCAPs clustering with multiple runs per subject

[eburkevt]

I'm trying to use iCAPs to assess co-activation patterns during a movie viewing during fMRI. Each participant has eight fMRI runs (watching eight different short movies) of about 5 minutes each (four runs on day 1, and four runs on day 2). Our TR = 0.900 seconds, 330 volumes per run.

I examined one subject at random and got very promising results after clustering with K=20. Eight of the 20 clusters appear to be canonical networks such as default mode, frontoparietal and salience networks.

Here is DMN (from eight runs, one participant)

Here is what appears to be the salience network

and here is what appears to be the frontoparietal network

However, when I tried 20 participants (8 runs each, so 160 runs total), the clustering results at K=20 yielded no canonical networks at all. All 20 clusters looked like noise or some other artifact. I looked at participants individually - most had canonical networks (to varying degree), but the when clustering over all runs/participants, the networks disappeared.

Is there anyone who understands iCAPs that might point me to what the issue is?

Thanks!

[eburkevt]

Resolved issue. The issue was with subject-specific gray matter (GM) masks. I used the probseg GM images output by fmriprep. These GM maps tend to be conservative. The iCAPs algorithm finds the intersection of GM masks, and due to the conservative nature of the subject-specific GM maps, tended to erode the final group GM mask (an intersection of all subject-specific GM maps) to only a small fraction of GM.

So, instead I went with an all-subjects conjuction GM mask for all subjects (one GM mask for all). It worked much, much better.