Some possible sources of differences between pTFCE and FSL's TFCE are:
1) differences in the T(Z)-score map due to the different statistical test applied (parametric test in SPM vs. permutation test in randomise)
2) accuracies during smoothness estimation (the SPM Toolbox uses SPM's smoothness estimates, see this link: https://github.com/spisakt/pTFCE/wiki/Some-important-notes-on-smoothness-estimation)
3) differences in the algorithms.
The effect of 1 and 2 will only be pronounced if there are serious non-Gaussianities in your data. The current, Gaussian Random Field-based implementation of pTFCE cannot be applied to largely non-Gaussian data, e.g. in VBM analysis. (Permutation-test based pTFCE is under development and will vastly extend the applicability of cluster-based belief boosting).
Algorithmic differences (3), as based on our simulation studies (DOI: 10.1016/j.neuroimage.2018.09.078), should only be pronounced in case of very large clusters. (See AUC nAFROC values on Figs 6 and 7). As the simulation suggests, in such cases, the original TFCE formulation seems to be overly optimistic (by falsely detecting some "background" voxels) while pTFCE tends to give a more pragmatic cluster enhancement, with a strict control of false positive voxels and, therefore, a better localisation performance.
Which of the above symptoms seem to fit to your data (Non-Gaussianity or very large activation clusters)?
I can help you decide if you share with me some data. E.g. the *_tfce_corrp_tstat* and the spmT* images and maybe your degrees of freedom.
University Hospital Essen, Essen, Germany