Fixelcfestats: nonstationarity

Orjan · May 15, 2016, 9:25pm

Hi,
I was trying fixelcfestats with the -nonstationary option and it seemed to have an “unrealistic” effect on the outcome with significant effects in many tracts but with little variation in pvalue. I am doing a group comparison (one sample ttest based on difference images calculated with fixelcalc; the data were preprocessed and analyzed according to the new documentation) and I know that there are global differences in FA - if that’s a clue - and I did the intensity normalization with a CSF mask. I would just like a bit more information on how the nonstationarity is calculated so that I can understand what is going on and if the results are correct.
Cheers,
Örjan

Dave · May 16, 2016, 12:25am

Hi Örjan,
As documented in the citations section of the help, the method implemented is:

ncbi.nlm.nih.gov

Adjusting the effect of nonstationarity in cluster-based and TFCE inference.

G Salimi-Khorshidi, SM Smith and TE Nichols, NeuroImage, Feb 2011 01

In nonstationary images, cluster inference depends on the local image smoothness, as clusters tend to be larger in smoother regions by chance alone. In order to correct the inference for such nonstationary, cluster sizes can be adjusted according to a local smoothness estimate. In this study, adjusted cluster sizes are used in a permutation-testing framework for both cluster-based and threshold-free cluster enhancement (TFCE) inference and tested on both simulated and real data. We find that TFCE inference is already fairly robust to nonstationarity in the data, while cluster-based inference requires an adjustment to ensure homogeneity. A group of possible multi-level adjustments are introduced and their results on simulated and real data are assessed using a new performance index. We also find that adjusting for local smoothness via a separate resampling procedure is more effective at removing nonstationarity than an adjustment via a random field theory based smoothness estimator.

I added this as an option since it was fairly trivial to implement. However, unfortunately I have not had time to test it thoroughly. In my little experience, it reduced the spatial extent of my significant results. However, I guess this was to be expected since the diseases involved fairly large tracts.

Ideally it would be nice to test it with some simulated pathology. As mentioned in the CFE discussion, it would also be nice to explore if there is any orientational sensitivity issues (like TBSS), and if there is, determine if non-stationarity correction fixes this. Unfortunately I can’t see anyone in our lab getting time to look at this any time soon.
Cheers,
Dave

Orjan · May 16, 2016, 5:59am

I should have seen that; I’ve always stopped scrolling at standard options My bad.
Sorry to waste your time.
Örjan