Hi Charis,

Sounds like your matlab stuff is working just fine.
The difference in means is due to grand mean scaling that occurs in FEAT.
This multiplies the whole 4D data by a single scaling factor in order to
make the mean value (over all 4D) equal to a constant value of 1000.
The reason this is done is so that different subjects can be compared - it \
doesn't affect any of the within-subject statistics though.

With regards to smoothing, it is a little confusing given that SPM uses
the HRF in two different ways.  Basically you are right in that in FSL
the HRF convolution is done to the model and not the data.  However, this
is *not* for the purposes of smoothing - it is to correctly model the link
between neural activity and blood flow.  Once this is done, and the initial
model is generated, any filtering (high-pass or low-pass) is done to
both the data and the model, since they are related by an equation and you
need to do the same processing to both sides in order to maintain the equality.
(Also, note that high-pass filtering is not smoothing, it is trend removal)

In SPM the confusing factor is that they use the HRF both as we do in FSL
*and* as a smoothing kernel.  Hence they apply it to the data in order to
smooth (low-pass filter) it.  So SPM spplies the HRF to both sides once
for the purposes of smoothing, after having applied it once to the model
(design matrix) for the purposes of physiological modelling.

I hope this makes things clear.
I'm not an expert on SPM implementation, so I can't really tell you more
than this, but that is the theory behind it.  There is no compelling reason
that the HRF should be used as a smoothing kernel - a Gaussian or any classic
low-pass filter would be just as good if the parameters are set sensibly.
What is important in SPM is that they do sufficient smoothing so that this
'colouring' of the data swamps the intrinsic smoothness (autocorrelation)
present in the data.  Since a general autocorrelation estimation technique
is used in FEAT (FILM) we don't need to do this, and can apply pre-whitening
which is the Best Linear Unbiased Estimator for the GLM, leading to more
sensitive statistics - hence my comments in the previous email.

All the best,
        Mark




Charis Tzagarakis wrote:
> Date sent:              Tue, 29 Oct 2002 14:52:49 +1100
> Send reply to:          FSL - FMRIB's Software Library <[log in to unmask]>
> From:                   Mark Jenkinson <[log in to unmask]>
> Subject:                Re: [FSL] FEAToutput, ANALYZE hdr and HRF parameters.
> To:                     [log in to unmask]
>
> Hi there!
>
> Mark,thank you very much for your reply/comments.
>
> When I look at the values in filtered_func_data and compare them to those of the
> original data, I notice that they differ by a factor of almost 1000 :I open the file in
> MATLAB.I also visualize it in that environment, so I can see that the images are not
> 'scrambled', and basically look the similar to those I feed in FSL -but they have better
> contrast, as they should!.I have tested my 'getANALYZE' routine independently and it
> seems to be reading the data correctly. Still this difference in mean puzzles me and I
> was wondering if it is something reasonable (ie simply the result of the filtering) or
> whether I should look more into it.
>
> Concerning the low-pass filtering,you  raised an issue that I would like to revisit,
> because I think I have probably misunderstood something here (in regards to both how
> FSL works and how it contrasts with SPM):
> I equate the process of convolving the data with the HRF to low-pass filtering.From
> what you mentioned in your posting I gather that this is not so.
> 'Taking it from the begginning' then, am I correct in the following statements ?:
> In FSL:
> Eessentially what happens is that the design matrix gets convolved with the HRF - and
> that, plus the high pass filteringon both sides of the GLM, is all the smoothing that takes
> place in the temporal domain.
> In SPM:
>  There is the added step of applying the low pass filter on both sides of the
> GLM(signal+design matrix),  what they call 'coloring',I suppose.
> If these statements are correct, I am a little confused about the HRF convolution step -
> isn't that in a certain sense a low pass filter in itself, and if that is the case, wouldn't
> SPM be 'filtering twice' the design matrix?
>
> Cheers !
> Charis.
>
> Charidimos Tzagarakis MD
> Brain Sciences Center and University of Minnesota.
> Address:
> Brain Sciences Center (11B)
> Minneapolis VA Medical Center
> One Veterans Drive
> 55417 Minneapolis MN
> USA
> Tel:612-725-2000 ext 1756
> email: [log in to unmask]