 |
 |
Hi Philipp Neither temporal concatenation nor full TICA are specifically designed for resting data. Temporal concatenation is useful on resting data as there are no further assumptions about temporal coherence made during model estimation. The temporal concatenation approach performs a rank-1 approximation to the full set of time courses as a post- processing step (i.e. unrelated to the model estimation!) as this is potentially a useful summary statistics on non-resting data. These modes are always meaningful in terms of the summary quantity calculated (it's the first principal eigenvector calculated from the set of all time courses) but likely not very interpretable when using resting FMRI data. We are currently evaluating new options to melodic specific to resting-FMRI which will permit a between subject/group analysis in a semi-parametric way, watch this space. In the meantime, the tXX.txt files do contain all the separate time courses as additional columns so you probably would want to derive your own summary stats that is sensitive to modulations of the resting patterns. hth Christian On 21 Nov 2008, at 14:29, Saemann Philipp wrote: > Hello Andreas, > > thanks once more. > > I think I better briefly describe what we did: > > We performed a multi-session temporal concatenation mode MELODICA > (v4.0) on > the original time-series on 93 epochs of each 5 minute resting scans. > This epochs come from 4 different sleep stages. The result in > principle > delivers typical resting networks (29, about 9 meaningful RSNs). > > We simply are stuck with the problem of how/if to correctly test the > subject modes (as contained in the box plots) of the four sleep stages > against each other. > > It seems that if negative mode values indicate strong > anticorrelation of > that specific individual, the modes are not distributed > parametrically from > low strength to high strength, but are mirrored round zero; so > maybe, the > absolutes of these values should be taken? > > So may be question could be re-formulated: Does time concatenated > PICA mode > of analysis that is recommended for resting state analysis deliver > meaningful subject mode values? > > Sorry for the longish correspondance here... > & thanks in advance, > Philipp > > > At 15:12 21.11.2008 +0100, Andreas Bartsch wrote: >> Hi Philipp, >> >> I think I got you wrong (or you got me wrong): I was not talking >> about any > F-testing, just about the multi-session/subject tensor-ICA on FMRI > time > series. Here, if you have one subject with an S-mode of 1 and > another with > -1 you may in fact assume that their time-courses OR > activation/deactivation maps behave the opposite way (i.e. are > anticorrelated, as you say). However, for resting state data you > cannot > assume that the FMRI time-series are temporally consistent across > subjects > whereas the powertransformed time-series may be consistent. Thus, > either > you run melodic in the full tensor-ICA mode on powertransformed data > or in > the multi-session temporal concatenation mode on the original time- > series. > The latter is recommended on the web, at least if you primarily want > to > identify the networks. >> Does that help? If not, I guess Christian will jump in... >> Cheers- >> Andreas >> >> >> ________________________________ >> >> Von: FSL - FMRIB's Software Library im Auftrag von Philipp G. Saemann >> Gesendet: Fr 21.11.2008 10:42 >> An: [log in to unmask] >> Betreff: [FSL] Meaning of S-modes in resting network group melodica >> >> >> >> Hello Andreas, >> >> thanks a lot for your answer on the F-test issue. >> >> If I understand it correctly, very low negative values would indicate >> strong "similarity" between the individual time course >> and the first eigenvariate time course, but in an anticorrelated >> manner. This >> would somewhat imply that the negative S-modes need to be >> flipped for testing against zero AND for comparison between >> subjects to come >> back to a parametric scale from "low strength" to "high strength" >> for resting >> analyses. >> >> (Take for e. g. the default mode network that is robustly detected in >> practically any group) >> >> My question now is: is the first eigenvariate (in time concatenated >> group >> analysis) constructed from somewhat time locked signal courses or >> from the >> unchanged mix of raw courses as they are in the group data? >> >> If not, then it seems at first sight that generation of the subject >> modes and >> their parametric value are somewhat not meaningful for resting data >> at all. >> >> So, e. g. correlation between network "strength" (from the mode >> values as >> they come from FSL) and a parametric behavioural measure seems >> obsolete to >> us. However, as we intended to do so, we are stuck here... >> >> Does anybody have an explanation of what the S-mode (across >> subjects of >> ONE component) as coming from FSL in the time-concatenated group mode >> really indicate? In published (users') work it says something like > "strength of >> BOLD fluctutations". >> >> Again thank you very much in advance for any hints on this, >> >> Philipp >> >> Max Planck Institute of Psychiatry >> NMR Research Group >> Munich >> >> > Max Planck Institute of Psychiatry > NMR Research Group > Kraepelinstr. 2-10 > 80804 Munich > Mail: [log in to unmask] > Phone: 0049-89-30622-413 _______________________________________________ Christian F. Beckmann, DPhil Senior Lecturer, Clinical Neuroscience Department Division of Neuroscience and Mental Health Imperial College London, Hammersmith Campus Rm 419, Burlington Danes Bldg, Du Cane Road, London W12 0NN, UK Tel.: +44 (0)20 7594 6685 --- Fax: +44 (0)20 7594 6548 Email: [log in to unmask] http://www.imperial.ac.uk/medicine/people/c.beckmann/ Senior Research Fellow, FMRIB Centre University of Oxford JR Hospital - Oxford OX3 9DU Tel.: +44 (0)1865 222551 --- Fax: +44 (0)1865 222717 Email: [log in to unmask] http://www.fmrib.ox.ac.uk/~beckmann