Dear SNR conversationalists,
 
If the goal is to say that "signal:noise is good enough/not good enough for purpose X", where purpose X is a frequentist hypothesis test, then it seems that ultimately a power analysis needs to be performed. Thus, as a reviewer, I would be unmoved by (or rather, I would find per se irrelevant) a signal:noise value obtained in a manner which does not correspond precisely to how test statistics are to be generated for the hypothesis test. Thus, if one intends on doing voxel-wise tests, computing SNR in ROIs is off the point. Likewise, SNR for one paradigm cannot be related to the signal:noise for another paradigm without making assumptions. The main points are that 1) SNR is not the ultimately relevant fact for hypothesis testing, power is and 2)while of course SNR plays a role in power computations, the SNR needs to be appropriate for the power calculation to make sense; the approriate SNR calculation would be one that corresponds to the test that one intends on performing.
 
Eric
 
 
 
----- Original Message -----
From: [log in to unmask] href="mailto:[log in to unmask]">Kingsley, Peter
To: [log in to unmask] href="mailto:[log in to unmask]">[log in to unmask]
Sent: Thursday, June 09, 2005 12:27 PM
Subject: Re: [SPM] Signal to noise ratio

Dear Bas, Marko, Amit, List:

These are some very good points.  Amit's original post implied measuring the SD over time, as recommended in Bas's reply.  It is important to distinguish between true (thermal) noise, and the actual variance (or SD) of data.  When pixels are averaged over space, other contributors to the variance include biological heterogeneity and possible image artifacts.  When averaged over time, the measured variance can be increased by physiological noise (blood flow, blood pressure changes, breathing), by subject motion (including breathing, swallowing, eye movements, blood flow), and by hardware instabilities (including gradient warming in long fMRI acquisitions, which can cause a slow drift in the slice locations).  So there are several questions to be asked before measuring the noise over time: Should you do motion correction first?  Should you do a linear (or other) detrending to reduce contributions from "scanner drift"?  Since the signal intensity changes by only a small amount with activation, choosing the signal from a rest or activation period will have minimal effect on the calculated SNR.  However, if both periods are used in calculating the noise, you have another factor contributing to the measured SD over time.

So, if you are trying to convince reviewers that you have sufficient SNR for your measurements, I suggest measuring SNR with the signal over a large ~homogeneous region, and the noise from a background measurement (with the correction factors mentioned in my earlier post).  If you are trying to determine whether an observed signal difference (e.g., task vs rest) is significant relative to the signal SD, I think you want the SD from the rest (or task) periods (after motion correction and high-pass filter).

As to what SNR levels are considered good, the easy answer is "it depends" (on pulse sequence, TE, TR, T1, T2, pixel size, +/- parallel imaging, etc.).  With proton-density SE non-EPI images (long TR, short TE) you can easily get SNR>100:1, especially in edema or CSF.  With diffusion tensor imaging (DTI), SNR~15-30 for brain tissue in the b=0 images is commonly reported, due to the long TE and the EPI acquisition (both of which are similar to fMRI).  If your SD ("noise") is measured over time or space so that it has non-noise contributions, your calculated SNR may be even lower.

Since I won't make it to HBM, I hope that any brilliant SNR insights from discussions there will be posted to the list.

--
Peter B. Kingsley, MRI Physicist        [log in to unmask]
Department of Radiology / MRI
North Shore University Hospital
300 Community Drive, Manhasset, NY 11030, USA
Tel (516) 562-2842       Fax (516) 562-3561



-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]On
Behalf Of S.F.W. Neggers
Sent: Thursday, June 09, 2005 3:49 AM
To: [log in to unmask]
Subject: Re: [SPM] Signal to noise ratio


Dear Amit, Marko, List,

I also have a (probably naive) question regarding the SNR measures described
by Peter Kingsley in his email (just re-read it) as well. When I understand
his post well, this classical type of SNR is for noise over voxels within a
single MR image. This might be interesting for anatomical images. For fMRI
though, I'd say it is much more interesting to now something about
singal-noise relation ships in time instead of in space, i.e. random
variations in time with respect to mean signal in time, or even better,
task-rest related amplitudes with respect to noise. To get a first idea on
SNR i sometimes simply divide mean signal (over time!) of a voxel or of the
average time signal of bunch of voxels in an ROI, and divide it through the
SD (over time!). One can then make an SNR image when doing this for each
voxel in a time series. For functional imaging that is much more informative
i'd say. There definitely are better measures for detectability of a signal
based on SNR measures at rest, I would also be interested in hearing about
them (and I'll dive deep into Medline of course). Perhaps we could explore
posters on fMRI-SNR measures next week in Toronto at the HBM together ;-)

Thanks for all your responses,

Bas

> Op donderdag 9 juni 2005 08:40, schreef Marko Wilke:
> > Hi Amit,
> >
> > > How can I calculate the signal to noise ratio in an fMRI data set? (ie
> > > make an SNR map) Is SNR contrast related (ie a different SNR for each
> > > contrast), or is it independent of the task (in which case I can just
> > > use a time series from a baseline rest condition)?
> > >
> > > Finally, what SNR levels are considered "good"?
> >
> > there was a very nice and very thorough mail on that from Peter Kingsley
> > on March, 24th, 2005, on this issue. I think most of your questions will
> > be answered there.
> > Best,
> > Marko

--
--------------------------------------------
Dr. S.F.W. Neggers
dept. of Psychonomics,Helmholtz Institute
Utrecht University
Heidelberglaan 2
3584 CS, Utrecht, room 17.09
the Netherlands
Tel: (+31) 30 253 4582 Fax: (+31) 30 2534511
E-mail: [log in to unmask]
Web: http://www.fss.uu.nl/psn/pionier
--------------------------------------------
_____________________________________________________________________

The information contained in this electronic e-mail transmission
and any attachments are intended only for the use of the individual
or entity to whom or to which it is addressed, and may contain
information that is privileged, confidential and exempt from
disclosure under applicable law. If the reader of this communication
is not the intended recipient, or the employee or agent responsible
for delivering this communication to the intended recipient, you
are hereby notified that any dissemination, distribution, copying
or disclosure of this communication and any attachment is strictly
prohibited. If you have received this transmission in error, please
notify the sender immediately by telephone and electronic mail,
and delete the original communication and any attachment from any
computer, server or other electronic recording or storage device
or medium. Receipt by anyone other than the intended recipient is
not a waiver of any attorney-client, physician-patient or other
privilege. Thank you.