Dear Katarzyna,

Cluster correction is quite difficult, this is not as easy to do as voxel-wise correction, because for cluster correction you need the infos about the image's topology, which is something totally subjective and calculated differently by various neuroimaging packaging. The key being the estimation of smoothness.

I myself tried to make a generic script to do that but I had to give up, as neuroimaging tools usually use a combination of:
1- an estimation of smoothness on the original image (usually based on the residuals after fitting a GLM model),
2- addition of the smoothness parameters used in preprocessing pipeline (so that this part does not require any estimation since the software knows internally the exact parameters of the smoothness it adds to the images, hence why most neuroimaging softwares prefer to use this approach).

In the end, I did not find a way, nor on the SPM mailing list or FSL or anywhere else, to reliably estimate the smoothness of just any neuroimage, because the neuroimaging smoothness estimation tools are not made to do that on an already processed image.

That said, there are two workarounds:
1. if you have an SPM.mat, containing the design which generated the images you want to cluster correct, then you can get the smoothness estimates from there and do your cluster correction (but then you can just open the SPM.mat and do the cluster correction in SPM directly, using the Results button).
2. it is possible to do a cluster extent size threshold, there is a script posted on the SPM mailing list that calculates the topology directly from the image using a k-neighbor kind of approach, but I can't find it right now (it's not the one by Tom Nichols). However, this approach was clearly highlighted in Eklund's paper ("Cluster failure etc.") as being the kind giving rise to the biggest amount of error rate above the nominal FWE rate, and thus is strongly disadvised.
3. lastly, Scott D. Slotnick, who is also the author of the paper "Cluster Success", a reply to Eklund's "Cluster Failure" paper, developed an alternative approach which can estimate smoothness directly. You can find the scripts and more information here: https://www2.bc.edu/sd-slotnick/scripts.htm

In the end, I would just advise to use the SPM.mat if it is available, or just fallback to voxel-wise correction, which does not require any smoothness estimation, and can thus be applied on any statistical map (both p-FWE and p-FDR can be used). I have a (unfinished but working) script in the works if you want to do voxel-wise correction.

Hope this helps,
Best regards,
Stephen