Thank you both for the input. What I'm interested in is defining regions of interest in individual subjects from a task run. Hence, the parcellated analysis does not seem the right fit for this application (unless I had individual parcels defined in some other fashion, e.g. from an individual resting-state parcellation). A permutation analysis (PALM) cannot be applied on a single subject's data, with only one run (unless I run a level 1 analysis many times shuffling conditions within the run?) - I'll take a look at it though for future reference. This leaves me with Matt's suggestion of a gradient-based approach. I remember this from the HCP course indeed, and I was very eager to try it on my data after seeing the beautiful results Matt was getting using gradients on task & myelin maps. When I tried running -cifti-gradient on my own contrast maps (this is not HCP data), the results were not very clear visually - lots of noise and clear boundaries difficult to detect. Perhaps I need a little more guidance on getting the gradient approach to work for ROI definition. Should I do much smoothing on the surface before detecting gradients? Is there a way in which you post-process the gradient map to make boundaries clearer? Maybe I should take a look at the methods published by the Petersen group for RS based parcellation using gradients? Matt, please let me know the pipeline you would suggest. Thanks for your help! - Julien
On Sat, Oct 10, 2015 at 3:08 AM Anderson M. Winkler <wink...@fmrib.ox.ac.uk> wrote: > Hi Julien, > > Indeed, as Matt said, you can use PALM, which does then a permutation test > (it is superior to Monte Carlo tests in that MC don't bypass the usual > assumptions of parametric tests). Of course, if MC tests were better, PALM > would surely be using them :-) > > Currently there are some difficulties with the CIFTI format, though. It > can read/write dtseries if the Workbench is installed, but the CIFTI > functions aren't stable and can crash easily. However, PALM can also work > with FreeSurfer formats and with GIFTI and NIFTI, so that you can convert > from CIFTI to one of these formats, run the test, then convert back to > CIFTI using the Workbench tools. > > PALM has the ability to take into account the family structure of the HCP > data, and one of the reasons why it was developed was precisely so that we > could use permutation tests with HCP. > > All the best, > > Anderson > > > > > > On 10 October 2015 at 02:16, Glasser, Matthew <glass...@wustl.edu> wrote: > >> If you are after area-wise activation in individual subjects, would you >> be willing to do a parcellated (i.e. area-wise) analysis where you could >> use something simple like FDR or Bonferroni (and get substantial SNR/power >> benefits)? If what you are actually after is to define regions of >> activation (instead of assessing the statistical significance of the >> activation in every grayordinate), I wouldn’t use statistical thresholding >> but instead would use gradients in the effect size map to identify most >> probably boundaries. >> >> In most cases for multiple comparison correction we are recommending >> people to use FSL’s PALM, but it is permutation-based. >> >> Peace, >> >> Matt. >> >> From: <hcp-users-boun...@humanconnectome.org> on behalf of Julien Dubois >> <jcrdub...@gmail.com> >> Date: Friday, October 9, 2015 at 4:53 PM >> To: "hcp-users@humanconnectome.org" <hcp-users@humanconnectome.org> >> Subject: [HCP-Users] Cluster-extent thresholding of CIFTI data >> >> Dear HCP developers, >> >> I'm wondering if you have come up with a good cluster-extent thresholding >> procedure and implemented it as part of the workbench yet. I have seen a >> couple of threads here >> <https://www.mail-archive.com/hcp-users%40humanconnectome.org/msg00893.html> >> and here >> <https://www.mail-archive.com/hcp-users%40humanconnectome.org/msg00993.html> >> re: this issue. >> >> One strategy would be to go the Monte Carlo way, considering the two >> surfaces and the volume separately (cf. Hagler et al 2006 >> <http://www.sciencedirect.com/science/article/pii/S1053811906007919> for >> the surface, which seems to be implemented in Freesurfer as mri_mcsim >> <https://surfer.nmr.mgh.harvard.edu/fswiki/BuildYourOwnMonteCarlo> and >> mri_glmfit-sim for the surface; and of course the very popular AlphaSim >> <http://afni.nimh.nih.gov/pub/dist/doc/program_help/AlphaSim.html>, >> superseded by 3dClustSim >> <http://afni.nimh.nih.gov/pub/dist/doc/program_help/3dClustSim.html>, in >> AFNI for the volume). >> >> Another way would be to go the Random Field Theory route. It's done in >> the volume in SPM with spm_uc_clusterFDR.m; I'm sure it has been adapted to >> surface data in some software suites (looks like the matlab code >> <http://www.math.mcgill.ca/keith/fmristat/toolbox/stat_threshold.m> in >> this old thread >> <https://mail.nmr.mgh.harvard.edu/pipermail/freesurfer/2007-August/005881.html> >> could work). >> >> Finally, FSL's randomize is another option that would play well with both >> surface & volume at the same time, and it seems to be the route you've >> taken. However (*correct me if I'm wrong*), this is not applicable to >> single-subject data with one run only, which is one of my interests. >> >> Please advise on the current status of development, and whether I need to >> "hack" my own cluster extent threshold correction based on MC or RFT for >> CIFTI data (if anyone knows any other good code resources to perform these >> analyses on surface/volume data, please share!). >> >> Thanks, >> - Julien >> >> -- >> Julien Dubois >> Postdoctoral Scholar >> California Institute of Technology, Pasadena, CA >> http://emotion.caltech.edu/~jdubois >> >> _______________________________________________ >> HCP-Users mailing list >> HCP-Users@humanconnectome.org >> http://lists.humanconnectome.org/mailman/listinfo/hcp-users >> >> _______________________________________________ >> HCP-Users mailing list >> HCP-Users@humanconnectome.org >> http://lists.humanconnectome.org/mailman/listinfo/hcp-users >> > > -- Julien Dubois Postdoctoral Scholar California Institute of Technology, Pasadena, CA http://emotion.caltech.edu/~jdubois _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users
