Include TFCE thresholding

[mriedel56]

https://www.fmrib.ox.ac.uk/datasets/techrep/tr08ss1/tr08ss1.pdf

[tsalo]

This would fall under the new correction module, but I don't know which methods it's applicable to. Do you know if it would work with all CBMA and IBMA methods?

[mriedel56]

I believe it should be applicable to all methods, as long as the image being evaluated has a range of values i.e., not binary.

[nicholst]

TFCE is a particular transformation of cluster information into a voxel-wise map. It can be regarded as 'meta analysis' of using all possible cluster-forming thresholds. However, to combine information over all of these thresholds there are 2 arbitrary tuning parameters, H & E, which have only been validated for Gaussian image data. Hence, it isn't clear how TFCE can be applied to CBMA, as the maps never have a Gaussian distribution. For IMBA, it should indeed be a trivial implementation exercise, though note that no parametric results are available and so permutation is required to obtain a threshold that controls the Familywise Error Rate.

[mriedel56]

FWIW, Simon Eickhoff's group has been using a TFCE Toolbox (https://zenodo.org/record/2641381#.XRO9bS2ZNo4) for meta-analyses (e.g., https://www.nature.com/articles/s41380-018-0122-5). They use 0.6 and 2 for H & E, respectively. How the 0.6 was defined differently than 0.5, Im not sure though.

[nicholst]

Not clear how they create a z-score to feed into TFCE. I suppose they can turn their uncorrected p-values into Z-scores. However, in my experience meta-analyses are not dying for power, so I'm surprised they've reached for TFCE. Anyway, I'll ask Simon how they're using it.

BTW, default H & E parameters are 2 and 0.5, respectively. In the TFCE paper, Random Field Theory suggests E should be 2/3, but since all the evaluations were for E=0.5 we left it as is. I suppose may Simon is use that, or maybe they tried to tune it.

[tsalo]

@nicholst did you ever get a chance to talk to Simon about how they're using TFCE?

[nicholst]

I've pinged him...

[tsalo]

@SBEickhoff was kind enough to respond via email. The key takeaways were (1) neither @SBEickhoff nor @nicholst is certain that the costs of TFCE (namely, decreased spatial specificity) outweigh the benefits (increased power) in meta-analytic data, but (2) TFCE can be applied to ALE. Two Python-based open source implementations that we could adapt are in mne-tools/mne-python and Mouse-Imaging-Centre/minc-stuffs.

@mriedel56 do you still want to add TFCE?

[SBEickhoff]

Dear Taylor and Cody

Just wanted to add something: In my experience, TFCE is actually not generally more sensitive than cFWE anyways, at least not when you add some extent-threshold to counter the higher number of spurious / false-positive clusters. So it could be an interesting addition if easy to implement, but I would not consider it a highly relevant topic

Best regards Simon

Am Di., 1. Sept. 2020 um 18:48 Uhr schrieb Taylor Salo < notifications@github.com>:

@SBEickhoff https://github.com/SBEickhoff was kind enough to respond via email. The key takeaways were (1) neither @SBEickhoff https://github.com/SBEickhoff nor @nicholst https://github.com/nicholst is certain that the costs of TFCE (namely, decreased spatial specificity) outweigh the benefits (increased power) in meta-analytic data, but (2) TFCE can be applied to ALE. Two Python-based open source implementations that we could adapt are in mne-tools/mne-python https://github.com/mne-tools/mne-python and Mouse-Imaging-Centre/minc-stuffs https://github.com/Mouse-Imaging-Centre/minc-stuffs/blob/0bfb834896e91a540774607ca7817ca8fcf41573/python/TFCE .

@mriedel56 https://github.com/mriedel56 do you still want to add TFCE?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/neurostuff/NiMARE/issues/145#issuecomment-684993329, or unsubscribe https://github.com/notifications/unsubscribe-auth/AE2Y32QQL2JKHU67ZRCOOSTSDUQW5ANCNFSM4H26GBDQ .

--

Univ.-Prof. Dr. med. Simon B. Eickhoff

Director, Institute of Systems Neuroscience Heinrich Heine University Düsseldorf 40225 Düsseldorf, Germany Tel: +49 211 586729-127 <0211%20586729127> eMail: Simon.Eickhoff@uni-duesseldorf.de http://www.uniklinik-duesseldorf.de/systemneurowiss

and

Director, Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour ) Research Centre Jülich 52425 Jülich, Germany Tel: +49 2461 61 1791 <02461%20611791> eMail: S.Eickhoff@fz-juelich.de http://www.fz-juelich.de/inm/inm-7/EN/Home/home_node.html