LLN, le 5/12/06
Dear David,
Below, I copied/pasted from the archives various answers by the
gurus. What you're looking for should be somewhere in these lines...
Note that some of these messages may apply only to previous versions
of SPM (unfortunately, I didn't copy the dates, but it should be easy
to search and find the original posts).
Hope this helps,
Mauro.
--------------------------------------------------------
How to determine High-pass filter cut-off point?
The cut frequency should be determined from the characteristics of
the noise rather than the characteristics of the signal. Since a
large part of the low frequency noise is system and sequence
dependent, the noise characteristics could be determined scanning a
phantom. When doing this it is very important to use a phantom with
internal structures and brain like T1 and T2* parameters. If you make
such an experiment and make power density plots of time series from
pixels near an edge (use for instance pyulear in matlab signal
processing toolbox) you will probably find something like this:
log(Power)
|
| \
| \ signal
| \ _/
| \ | | noise
| \______ |_|_______/____
| | |
| | |
| | |
-----------------------> log(Frequency)
| | |
f-cut f-paradigm f-nyquist
the slope of the low frequency noise will probably be between 1/f^2
to 1/f and f-cut around 1/100s=0.01Hz. In this case the simple
formula would read T=100s. The signal of the paradigm (event related
or box-car) is most easily detected when it is located in the white
noise region. But this has to be assured when you set up the
experimental design.
Torben E. Lund Danish Research Centre for MR - email:
[log in to unmask]
-----------------------------
The choice of the highpass cut-off would ideally maximise the
signal/noise ratio. However, we cannot distinguish signal from noise
on the basis of the power spectrum of the data alone. One choice of
cut-off is to minimise the loss of signal, the frequency components
of which are inherent in the design matrix X (look at the power
spectrum produced in the SPM graphics window). SPM99 offers such a
default cut-off based on twice the maximum SOA between the most
frequently occurring condition. If the real signal corresponds to the
design matrix, there's no gain increasing the cut-off, though one
might gain degrees of freedom (by filtering less, ie with fewer DCT
components in the filter), this gain could be offset by additional
white noise passed. More importantly, you might lose sensitivity if
your cut-off started to allow additional nonwhite noise. If the
default cut-off period based on X is too great, the gain in signal
passed can be outweighed by the extra noise passed. Experimental
designs should therefore not embody significant power at low
frequencies, i.e, conditions to be contrasted should not live too far
apart in time. The cutoff of 128 that is now the default in SPM2 is
based on minimising noise alone (given the rough generalisation that
low-frequency noise tends to increase above this time-contant, though
this will depend on many factors, eg scanner, subject, etc), in order
to minimise the risk of people going with defaults based on X that
are too high based on the noise.
Rik Henson
------------------------------
If the design is alternating with 25 seconds "rest" and 25 seconds
"stim" then this is roughly a square wave with period 50 seconds. If
the system which transforms neural activity change to BOLD fMRI
signal change is linear, then there will be no experimental effect in
the fMRI signal below (1/50) Hz. Now, given that fMRI time series are
discrete, it is best to first think of the cutoff in terms of cycles
per total sampling period. So, for example, if your experiment was
100 seconds long, then the most stringent hi-pass filter you could
use without eating into your task effect would correspond to 1
cycle/(total sampling period), or (1/100) Hz. If instead your
experiment was 300 seconds long, then the most stringent hi-pass
filter you could use without eating into your task effect would be 5
cycles/(total sampling period), or (5/300)=(1/60) Hz. So, the answer
to your question is given, I believe, in general by the following
formula:
given a two condition blocked design with fundamental period T
seconds (e.g., T=50 seconds in your case) and an experiment of
duration N seconds (and assuming T goes evenly into N), then
most stringent hi-pass cutoff (in Hz) = ((N/T) - 1) / N
(Note that this formula does not involve the TR.)
Eric Zarahn
------------------------------
Why is "128" the default value of Highpass filter?
The (default suggestion) for the HP threshold in SPM2 is hardcoded to
128 seconds. It used to be determined on the basis of the
experimental design (in SPM99) and is now "determined" on the basis
of the "known" behaviour of low-frequency noise. If you specify a
"too short" HP cut-off period compared to your experimental design
you will filter away "interesting" experimentally induced variance.
The rule of thumb there is still:
2*(the longest time between onset of consecutive epochs of the same type).
I suspect that the changed default suggestion reflects a strong
recommendation to design your experiments such that that rule of
thumb gives you a value <=128 seconds. If your experimental design
forces you to use a longer period you will remove less of the
low-frequency noise and that in turn will affect SPMs'ability to
estimate the temporal autocorrelations.
Jesper Anderson
------------------------------
>Dear SPMers,
>
>I am running a blocked design with 2 conditions, repeated twice for
>a total of 4 runs, with 2 different conditions.
>I am currently using the High-pass filer cutoff default in SPM5 of
>128s. Does this seem resonable, or is ther a more precise way I can
>determine what value to use? My TR is 2.24 seconds.
>
>Thank you.
>
>David Kideckel
>PhD Student, University of Toronto
>Canada
--
_____________________________________
Help fighting hunger: http://www.hungersite.com
Just click your mouse and sponsors of The Hunger Site donate a
serving of food to a person in need - at no cost to you.
______________________________________
Mauro PESENTI
Research Associate, National Fund for Scientific Research (Belgium)
Unite de Neurosciences Cognitives
Departement de Psychologie
Universite Catholique de Louvain
Place Cardinal Mercier, 10
B-1348 Louvain-la-Neuve
tel.: +32 (0)10 47 88 22
fax: +32 (0)10 47 37 74
E-mail: [log in to unmask]
http://www.nesc.ucl.ac.be
http://www.nesc.ucl.ac.be/mp/pesentiHomepage.htm
|
