Dear Saaussan,
Just a few comments (again authoritative control will be appreciated) :
Here is my understanding to noise in fmri experiments. Please correct me if
i'm
wrong.
If you take an roi (could be one voxel) under any condition and plot the
density
power spectrum using the autocorrelation function, you will get the "1/f"
graph
and possibly some "squiggles".
The "1/f" graph is the so called "DC" component of the spectrum while the
"squiggles" are the "ac" components.
I am not kind on auto-correllogram, and I don't know what "squiggles" are
(may I guess that they represent some small signs of periodicity as small
waves on each side of the auto-correlogram - thanks to correct me). However
I wouldn't qualify 1/f noise as DC since the DC reefers to the 0 frequency
(== decay from zero).
1/f noise do have a clear higher power spectra in low frequencies with a
reduction of it proportional to 1/f. Most of the higher part of the spectra
is flat == white noise (although I remember a abstract mentioning some
higher frequency => this could depend of your TR).
I think that you may found a better explanation by reading this chapter
(part 3)
http://www.fil.ion.bpmf.ac.uk/spm/course/notes97/Ch9.pdf
A better view for the "ac" components may be obtained by removing the "DC"
component, which hides lower "ac" freqeuncy "squiggles". [Simply subtract
the
mean of spectral density]
In our lab, we reduce the noise by applying a bandpass filter (hamming for
example), with the cuttoff frequencies based on the experimental design.
Example1: For a block design 30secON/30secOFF, we use (4*30sec, 30sec) as
cuttof
frequencies.
Example 2: For an experimental design which has 30secON/30secOFF &
15secON/15secOFF, we use (4*30sec,15sec) cutoff frequencies.
I would appreciate another input on that point, but it seems to me that 15
and 30 sec high pass filtering is much too large. Sure that if you filter
at quite the paradigm frequency you will get something. But how do you
build your statistic after that tremendous loss of independence ? Not
taking this lost of degree of freedom into account may give you a too
permissive test.
Another point is that you will loose some interesting temporal evolution of
the signal.
Our statistics improves (for our data where physiologic noise density is
comparable to our signal) by applying a bandpass filter compared to
modeling the
the "i/f" noise. However, we didn't test for zero mean, and guassian
distribution
of the noise after applying the bandpass filter.
This is very close to what is doing SPM => it removes low frequency noise
with appropriate covariates, and high frequencies by convolving the signal
with an 'hemodynamic response' like function (something also important for
statistical inference since it serves as an evaluation of loss of
independence between each observation).
-s madi
Drexel university
This is very much apart, but I have looked in all my atlas, and haven't
found Drexel, could you provide me some help ?
I hope that I haven't missed your question and that it could be of some
help
Best regards
Jack
________________________________________________________________
| Jack Foucher Universite Louis Pasteur |
| Institut de Physique Biologique UPRES-A 7004 du CNRS |
| 4 rue Kirschleger Tel: 33 (0)3 88 77 89 90 |
| 67085 STRASBOURG Fax: 33 (0)3 88 37 14 97 |
| France |
| Faster E-mail: [log in to unmask] |
| Other [log in to unmask] |
|_______________________________________________________________ |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|