Extremely large clusters with TFCE using FSL randomise

[aftonnelson]

Dear Neurostars community,

I have a question regarding the cluster sizes when using FSL randomise with TFCE.

I get very large clusters when applying TFCE (~35,000, see screenshot), which does not appear when applying, for example, voxel-based correction (~2,300, screenshot). I have attached screenshots of the tstat and pstat (red/yellow) map from the randomise with TFCE and the pstat from the randomise with voxel correction (black). There are also the cluster tables (TFCE-left, voxel-right) showing the cluster sizes.

I understand that TFCE is more robust and would likely lend to greater clusters sizes. However, the result seems strange that it would be so large (i.e., nearly 15x larger).

Is this indeed a strange outcome, or might it be plausible? Are there any other checks I can run?

For reference, the command from the log file:

TFCE: randomise -i filtered_func_data -o /stats -m mask -d design.mat -t design.con -e design.grp -n 500 -T --film

voxel: randomise -i filtered_func_data -o /stats -m mask -d design.mat -t design.con -e design.grp -n 500 -x --film

I get similar cluster sizes (or even larger) when I run it with more permutations.

[psadil]

This might not be too surprising. There is a phenomenon called “cluster leakage”. It’s not necessarily a problem, depending on your goals; the presence of a significant cluster only allows rejection that there are zero active voxels within that cluster (e.g., maybe only one active voxel), and if there is at least one voxel active within a cluster, then there is also at least one active voxel when that cluster is expanded. There is some belief that TFCE may be prone to cluster leakage (I don’t have a reference, unfortunately), so leakage could be at work here.

Again, whether it’s a problem depends on the kind of inferences you want to make. Regardless, you could also try the updated version method, probabilistic TFCE (Probabilistic Threshold-free Cluster Enhancement | pTFCE, Spisák et al (2019)), which should be less prone to leakage (that paper discusses the phenomenon).