 |
 |
Hi Tom, Thanks for your suggestions! I'm not interested in differences in sessions (since we only have one session) . The goal of our study is to find pain responding areas in the brain. We already did this with a classical GLM analysis. The reason why want to use PICA is to possibly find additional areas responding to pain in another way than the characteristic BOLD response(e.g. which would not been found with a HRF-Model). So my idea was to timelock, all IC-timecourses and decide on the basis of a statistical test (Hotellings T2), whether the timelocked (average) IC timecourse follows a regular run. And there my only hypothesis comes into the analysis: timelocked timecourses, which are not depending on the stimulus, when averaged, should null-out (random distribution around zero). IC-timecourses which show a regular run, when time- locked, are related to the stimulus. (like an ERP in EEG). I tried this with single subject PICA in my diploma thesis and I got some reasonable results. But I'm not sure if I can take this approach in a concatenated PICA. In the single subject PICA, the next step would be to run a PCA on the sorted (event related) IC-maps. Like this one should find ICs, which are similar in the spatial characteristics of the activations, as well as the event related timecourses. One could put such similar ICs in a random effects analysis and obtain common event related areas.... Bing et al. are presenting an interesting approach to sort ICs: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4351419 -Andreas Am 13.03.2008 um 10:46 schrieb Tom Johnstone: > Actually, I had a further thought. Would it be possible to calculate > the power spectrum of each session time series and feed that into > tensor ICA? Because the power spectrum ignores phase information, the > spectra of different sessions with different jittered timing would be > very similar. Would that be possible Steve, Christian? > > -Tom > > On Thu, Mar 13, 2008 at 9:44 AM, Tom Johnstone > <[log in to unmask]> wrote: >> Hi Andreas, Steve, >> >> I'm actually doing exactly the same type of analysis at the moment, >> and coming across the same sort of analytic questions (but not with >> pain stimuli). Actually, in our case we have each subject doing 3 >> scan >> runs/sessions, each with the same basic task but with differently >> jittered stimulus presentation. The sessions differ with respect >> to an >> external variable of interest. So we'd like to use ICA to find i) >> components that are common across all sessions and ii) components >> that >> differ between sessions. And for each of these, a) components that >> correspond to stimulus presentation and b) components that do not. >> >> So as I said, what we're doing is very similar to what you're doing. >> Previously Christian had given me some advice with this, which was >> indeed to use the multi-session concatenated ICA approach. Indeed, >> this gives separate estimates of component time series for each >> session, which can then be compared with the model using >> 'offline' GLM >> (because one needs to a different time series model for each >> session), >> allowing us to identify a) and b) as described above. >> >> It's the issue of identifying common and different components across >> sessions that's more tricky. The reason being that we need to >> identify >> which components in, say, sessions 2 and 3 correspond to a given >> component in session 1. One way to do that is by visual inspection. >> Alternatively, and somewhat intuitively, the way I've been thinking >> about this is that the time series will not be directly comparable >> across sessions, but the spatial structure of the components should >> be. So it should be possible to estimate the spatial covariance >> between pairs of components, and identify those with high >> covariance. >> Of course there's still the chance that the time series of spatially >> similar components might be very different, in which case they >> would >> presumably not represent functionally similar networks. Perhaps a >> further constraint could look for similarity in the frequency >> spectrum >> of spatially similar components. Frequency spectra would identify >> components with similar time series structure, but different >> absolute >> timing (i.e. phase). >> >> Then, if a component in one session had "matching" components in >> other >> sessions, it could be designated common to all sessions. If no >> matching components were found, it would be a session-unique >> component. >> >> Anyway - a bit of a ramble, but maybe something in there is of some >> use to you, or generates some further ideas. >> >> Cheers, >> Tom >> >> >> >> On Thu, Mar 13, 2008 at 9:10 AM, Andreas Pedroni >> <[log in to unmask]> wrote: >>> Hi Steve, >>> >>> Unfortunately, stimuli were administerd with a random jitter. So >>> tensor ICA would not work, I guess. >>> We would like to remain fully exploratory throughout the whole >>> analysis, so we would like to avoid a timeseries model with a >>> priori >>> hypotheses about hrfs. >>> Instead we want to try to timelock neural responses of the >>> events and >>> sort IC-timecourses, which are event related (showing an average >>> response (e.g. are not distributed around zero, in each timepoint) >>> from IC timecourses, which are not due to events (e.g. which shoul >>> distribute around zero, when averaged). >>> As mentioned before I tried this sorting procedure with single >>> subject PICA and it seemed to work out well, but I'm not really >>> sure, >>> about a concatenated PICA. Do the columns in the t**.txt files >>> represent the IC-timecourse for each subject? Finally, do we get >>> >>> many more ICs with 20 subjects than with 10? Or does it make >>> sense to >>> constrain the analysis to a certain number of ICs? >>> >>> Cheers! >>> >>> Andreas >>> >>> >>> >>> Am 13.03.2008 um 09:51 schrieb Steve Smith: >>> >>> >>> >>>> Hi - if you have the same timeseries model for all subjects then >>>> the easiest thing to do is to use Tensor-ICA (not concatenation) >>>> and enter the timeseries model in the MELODIC GUI in the same way >>>> that you do for single-subjects PICA analysis. >>>> >>>> Cheers, Steve. >>>> >>>> >>>> On 11 Mar 2008, at 09:51, Andreas Pedroni wrote: >>>> >>>>> Hi >>>>> >>>>> We are currently working on a study, where we try to identify >>>>> neural >>>>> structures responding to tooth pain. We already analyzed data >>>>> with a >>>>> classical GLM. Now we try to run MELODIC to possibly find other >>>>> neural >>>>> responses to administerd tooth pain. Our idea is, to rund a group >>>>> PICA >>>>> (concatenated subjects). In a second step we try to identify pain >>>>> related >>>>> ICs by timelocking IC timecourses and testing for each IC >>>>> timecourse, if >>>>> the time locked timecourse follows a regular run (e.g. IC >>>>> timecourses which >>>>> are not due to the pain stimulus should distribute around zero, >>>>> when time >>>>> locked to events, like a ERP). I already tried this with single >>>>> subject PICA >>>>> of data of an event related design and my results seemed to make >>>>> sense, >>>>> regarding the timecourses and the structures found. Now, my >>>>> questions are: >>>>> Can we do the same thing with the concatenated version of PICA? Is >>>>> it true, >>>>> that in the t**.txt files each column represents the timecourse >>>>> of a >>>>> subject? Finally, do we get many more ICs with 20 subjects than >>>>> with 10? Or >>>>> does it make sense to constrain the analysis to a certain number >>>>> of ICs? >>>>> >>>>> Thanks for any hints! >>>>> >>>>> Cheers >>>>> >>>>> Andreas Pedroni >>>>> Instiute of Psychology Zurich, Neuropsychology >>>>> >>>>> >>>> >>>> >>>> ------------------------------------------------------------------- >>>> --- >>>> ----- >>>> Stephen M. Smith, Professor of Biomedical Engineering >>>> Associate Director, Oxford University FMRIB Centre >>>> >>>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK >>>> +44 (0) 1865 222726 (fax 222717) >>>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve >>>> ------------------------------------------------------------------- >>>> --- >>>> ----- >>> >> >> >> >> -- >> School of Psychology and CLS >> University of Reading >> 3 Earley Gate, Whiteknights >> Reading RG6 6AL, UK >> Ph. +44 (0)118 378 7530 >> [log in to unmask] >> http://www.personal.reading.ac.uk/~sxs07itj/index.html >> http://beclab.org.uk/ >>