 |
 |
Hello professor Christian, Thanks a lot for that explanation, I really appreciate it. I already read the article that you told me, thanks but I still need some help on this, please. I would like to know if what I am doing is correct, if not I would greatly appreciate if you could correct me. My guess is that I have to apply the following steps: 1) To convert the tstast maps from randomise to zmaps we can use the following commands, righ?: fslmaths dr_stage3_ic00*_tstat1.nii.gz -mul 0 -add 1 grot.nii ttoz grot.nii dr_stage3_ic00*_tstat1.nii.g <dof> Where DOFs are the number of the subjects included in the analysis minus the number of explanatory variable (EVs) in my design, Is that right? 2) After that we would get the zstat1 map and I know I have apply GGMM using the second type of usage of melodic (getting the empirical null distribution) but I am not sure how to set up melodic. I was thinking we would run melodic as I described below: melodic –i bg_image (dummy option) --ICs=4d_zmaps_from_randomise_tstat1_or_tstat2 --mix=melodic_mix_33X32.txt (dummy mixing matrix) -o outputname -v --report –Ostats If not, Could you PLEASE explain to me how to set up melodic? Do I have to use a design matrix and a design contrast options in melodic, if so What are the designs and the "tstat images" (4d_zmaps_from_randomise: tstat1_and_tstat2 or tstat3 and Tstat4) that I have to use? 3) Once when we have the empirical null distribution (the zstat1 maps thresholded) I am not sure how to calculate the local FDR. Can we apply the following command? fslmaths thresh_zstat1.nii -ztop pmap1.nii obtaining the uncorrected p images, which can be use as a input in a “standard” FDR, thresholding at p<0.05, right?? As I am showing below: fslmaths grot_vox_p_zstat1 -mul -1 -add 1 1_minus_grot_vox_p_zstat1 fdr -i 1_minus_grot_vox_p_zstat1 -m mask -q 0.05 Thanks a lot for being so patient, I would be most grateful for any advice you could give on this matter of deep concern to me Thank you so much in advance Lorena --- On Wed, 3/16/11, Christian F. Beckmann <[log in to unmask]> wrote: > From: Christian F. Beckmann <[log in to unmask]> > Subject: Re: [FSL] GGMM > To: [log in to unmask] > Date: Wednesday, March 16, 2011, 4:56 PM > Hi Lorena, > > Sorry for the confusion - I'll try to be as clear > while keeping this concise: > > (i) The Filippini et al paper description is correct. The > dual regression output was thresholded using "local FDR". > This refers to the process of (a) estimating an empirical > null distribution, in our case based on the standard melodic > GGMM fit and (b) using this empirical null (rather than the > typically used assumed null distribution) within a standard > FDR thresholding procedure. See Efron, B Journal of the > American Statistical Association March 2004, Vol. 99, No. > 465, pp 96-104 for more info on how to do this in practice > > (ii) Unlike what's described in the Efron paper (estimating > the empirical null based on a spline fit with manual > intervention) I'm using the GGMM to identify the > empirical null as it's a very robust model and gives me > something fully automated > > (iii) While this will probably become the default > thresholding technique in melodic in future versions, it is > not yet a publicly available command line option in the > current release version. The first stage (fitting the GGMM) > is readily available (see the email you were referring to, > where I refer to the second type of running melodic) but > this will threshold at p>0.5 based on a full alternative > hypothesis test, i.e. not using local FDR. > > hth > Christian > > > > > > On 16 Mar 2011, at 21:05, Lorena Jimenez-Castro wrote: > > > Hello Eugene and FSL experts, > > > > I am sorry that I insist but in the Filippini paper, > specifically in the Materials and Methods section you can > read the following: > > > > Filippini et. al. > > "(ii) using these time-course matrices in a linear > model fit (temporal regression) against the associated fMRI > data set to estimate subject-specific spatial maps. Finally, > the different component maps are collected across subjects > into single 4D files (1 per original ICA map, with the > fourth dimension being subject identification) and tested > voxel-wise for statistically significant differences between > groups using nonparametric permutation testing (5,000 > permutations) (55). This results in spatial maps > characterizing the between-subject/group differences. > > > > These maps were thresholded using an alternative > hypothesis test based on fitting a Gaussian/gamma mixture > model to the distribution of voxel intensities within > spatial maps (see ref. 31 for further details) and > controlling the local false-discovery rate at P < 0.05" > > > > Now I am confused because after read that description > my understanding is that after he used the dual regression > analysis followed by randomise, he > thresholded the statistical maps from randomise using the > gaussian/gamma mixture model. I believe that also since > professor Christian F. Beckmann agreed with that in the post > previously mentioned (https://www.jiscmail.ac.uk/cgi-bin/webadmin > A2=ind1009&L=FSL&P=R17502&1=FSL&9=A&I=3&J=on&X=565E91456B012C3733&Y=lojicas%40yahoo.com&d=No+Match%3BMatch%3BMatches&z=4) > > > > That is why I greatly appreciate if somebody would > please help me with my previous questions or clarify this > concern to me > > > > Thank you so much in advance > > > > Lorena > > > > > > From: Eugene Duff <[log in to unmask]> > > Subject: Re: [FSL] GGMM > > To: [log in to unmask] > > Date: Wednesday, March 16, 2011, 1:53 PM > > > > > > Hi Lorena - > > > > > > On 16 March 2011 17:56, Lorena Jimenez <[log in to unmask]> > wrote: > > Dear FSL experts, > > > > We ran MELODIC on our resting state data using > multi-session temporal concatenation followed by dual > regression analysis including randomise. > > We had conducted an unpaired t-test in randomise > comparing two groups (16 controls versus 16 patients). > > Contrasts: > > 1.000000e+00 -1.000000e+00 > > -1.000000e+00 1.000000e+00 > > 1.000000e+00 0.000000e+00 > > 0.000000e+00 1.000000e+00 > > > > Now we would like to threshold the statistics maps > from randomise using Gaussian gamma mixture model as was > done by Nicola Filippini “Distinct patterns of brain > activity in young carriers of the APOE-��4 allele” > > I have being trying to figure out how to apply that > threshold but I could not find a solution and the only one > post that I found in the archives on this topic does not > have answer for its last question. (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1009&L=FSL&P=R17502&1=FSL&9=A&I=3&J=on&X=565E91456B012C3733&Y=lojicas%40yahoo.com&d=No+Match%3BMatch%3BMatches&z=4) > > > > > > > > There seems to be confusion on the method of > thresholding used in that paper (it is unclear in the > paper). GGMM was NOT used to threshold the > dual-regression results, only to represent the original IC > maps (where thresholding is primarily for interpretation as > these maps are not testing an experimental > hypothesis). Standard randomise voxelwise thresholding > was used. I don't think TFCE was used, but this is an > option. GGMM doesn't really make sense as a formal > statistical test in this context. > > > > Best, > > > > Eugene > > > > > > My guess is that I have to apply the following steps: > > > > 1) To convert the tstast maps from randomise to zmaps > we can use the following commands righ?: > > > > fslmaths dr_stage3_ic00*_tstat1.nii.gz -mul 0 > -add 1 grot.nii > > ttoz grot.nii dr_stage3_ic00*_tstat1.nii.g > <dof> > > > > Where DOFs are the number of the subjects included in > the analysis minus the number of explanatory variable > (EVs) in my design, Is that right? > > > > > > 2) After that we would get the zstat1 map and I know I > have apply GGMM using the second type of usage of melodic > but I am not sure how to set up melodic. I was thinking we > would run melodic as I described below: > > > > melodic –i bg_image (dummy option) > --ICs=4d_zmaps_from_randomise_tstat1_or_tstat2 > --mix=melodic_mix_33X32.txt (dummy mixing matrix) -o > outputname -v --report –Ostats > > > > If not, Could you please explain to me how to set up > melodic? > > Do I have to use a design matrix and a design > contrast options in melodic, if so What are the designs and > the "tstat images" (4d_zmaps_from_randomise: > tstat1_and_tstat2 or tstat3 and Tstat4) that I have to use? > > > > > > 3) once when we have the zstat1 maps thresholded, Can > we apply the following command? > > > > fslmaths thresh_zstat1.nii -ztop pmap1.nii > > obtaining the uncorrected p images, which can be > use as a input in FDR, right??? > > > > > > I know how valuable is your time, I really appreciate > any hint on this > > > > > > Thank you so much in advance > > > > Lorena >