A couple points. It's important to keep in mind that temporal ICA (TICA)
and spatial ICA (SICA) are not the same (in contrast to PCA where it doesn't
matter). For SICA the images are mostly indepenent, whereas the timecourses
can be correlated sometimes up to 0.3 or 0.4, vice versa for TICA. So there
are probably certain cases where one will work better than the other. That
being said, SICA is more suitable in general for fMRI for several practical
reasons I think. First, though it is still possible to estimate TICA
components, even with a hundred timepoints, the estimate is probably not as
robust since we have only a few hundred timepoints typically (this is the
point Steve is making below). Secondly, to do TICA you need to computer a
covariance matrix which is voxels * voxels...too large to do in practice, so
typically TICA is limited to one slice or a subregion. Some folks prefer to
use SICA first to identify regions of interest, then do TICA within only
these voxels. Finally, I think it makes sense to identify spatially
distinct regions which have coherent activity in time. There are more
things that can be said here, but I think I'll stop. ;-)
Hope this is helpful.
Regards,
Vince
> -----Original Message-----
> From: FSL - FMRIB's Software Library
> [mailto:[log in to unmask]] On Behalf Of Steve Smith
> Sent: Friday, February 23, 2007 12:19 AM
> To: [log in to unmask]
> Subject: Re: [FSL] TICA and SICA
>
> Hi,
>
> On 22 Feb 2007, at 16:26, pengxu wei wrote:
>
> > Dear Prof. Smith,
> >
> > Thank you for your correction. I had thought Melodic as a TICA
> > software for a while and now failed to traced the reason.
> >
> > It is said in temporal ICA the interpretation is that a set of
> > spatial patterns are activated by independent temporal processes.
> > But in spatial ICA, spatially independent brain sources or
> > components are calculated from fMRI data, of course the associated
> > time courses as well. So I think although both SICA and TICA will
> > get both spacial maps and associated time courses, the points are
> > different,e.g, which one is the first, spacial maps or temporal
> > processes?
>
> That's not really the point - the point is that unless you have an
> unusually high number of timepoints, you can't use temporal ICA well
> because you don't have enough timepoints to estimate the PDF across
> time.
>
> > And for SICA and TICA, the models used are different, it is said
> > the model that SICA used is for sparse distributed activation
> > paradigms. Vince said for most task-related components, the
> > results from both SICA and TICA are comsistent in spite of
> > different algorithms used. But to some components, like
> heart beat,
> > SICA cannot separate it well. The problem is, that paper was
> > publised more than 3 years ago so things may have changed. That is
> > why I asked such a question.
>
> We've not come across examples where the spatial independence model
> does not work well - MELODIC in my experience does a very
> good job of
> separating out physiological artefacts including breathing and
> cardiac related.
>
> Cheers.
>
>
> >
> > Best regards,
> >
> > wei
> >
> > Steve Smith <[log in to unmask]> wrote:
> > Hi,
> >
> > On 20 Feb 2007, at 02:44, pengxu wei wrote:
> >
> > > Dear all,
> > >
> > > In a paper (Vince D. Calhoun, T¨šlay Adali, Lars Kai Hansen, Jan
> > > Larsen, James J. Pekar. ICA of Functional MRI Data: An Overview.
> > > Fourth International Symposium on Independent Component Analysis
> > > and Blind Source Separation. 2003.281-288), it is said
> > > (because of both SICA and TICA have their own limits) ' In
> > > general it is recommended to
> > > explore the full spectrum of independency and perform both SICA
> > > and TICA'
> > > so we can get thorough results.
> > >
> > > After several years, now has Melodic, as a TICA method and with
> > > many revised versions, solved such limits such as ' the on/off
> > > structure of the
> > > activation time course doesn't match well with the sparse source
> > > assumption of the Bell-Sejnowski approach' so we don't need a
> > > additional SICA?
> >
> > MELODIC is spatial ICA ("SICA") not temporal, because FMRI data are
> > not well-suited to TICA - primarily because you don't have enough
> > timepoints to generate an accurate temporal PDF.
> > I can't think of any reason why you would particularly want to run
> > temporal-ICA on FMRI data as opposed to spatial-ICA.
> >
> > Cheers, Steve.
> >
> >
> > >
> > > Best regards,
> > >
> > > wei
> > >
> > >
> > >
> > > Finding fabulous fares is fun.
> > > Let Yahoo! FareChase search your favorite travel sites to find
> > > flight and hotel bargains.
> >
> >
> >
> --------------------------------------------------------------
> --------
> > --
> > ---
> > 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
> >
> --------------------------------------------------------------
> --------
> > --
> > ---
> >
> >
> > Sucker-punch spam with award-winning protection.
> > Try the free Yahoo! Mail Beta.
>
>
> --------------------------------------------------------------
> ----------
> ---
> 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
> --------------------------------------------------------------
> ----------
> ---
|