Hi Jeremy,

it really depends on the intrinsic SNR of your images. For example, with 
3T, 1mm isotropic voxels and a 12 channel head coil we have found that the 
decreased contrast from interpolating can reduce the test-retest 
reliability when using two mp-rages as opposed to one. Averaging will 
decrease the noise, but also decreases contrast due to the interpolation, 
so the effects on the CNR aren't always clear. For hires acquisitions (e.g. 
0.5mm) the averaging is needed because the intrinsic SNR is so low (8 times 
lower than 1mm), but with modern phased arrays and 3T images averaging 
multiple acquisitions isn't always a win. Particularly if you are imaging 
populations that move. Unfortunately CNR is difficult to measure well as it 
relates to structural analysis, since the performance of the algorithms are 
usually nonlinear in the CNR. That is, the CNR just needs to be good 
enough, after which classification accuracy asymptotes pretty quickly. You 
also have to ask CNR between what and what? Optimizing CNR between 
neocortical gray matter and the underlying white matter won't be optimal 
e.g. for distinguishing putamen from pallidum. And even if you pick two 
tissue classes, optimizing CNR won't necessarily optimize the sharpness of 
the boundary (which will blur with multiple acquisitions), potentially 
resulting in lower reliability even with increased CNR.

Sorry for the complicated answer. I think the best strategy is to look at 
your test-retest reliability for 1, 2 and 3 acquisitions, and see what is 
optimal. Multiple acquisitions can also be useful even if you find using 
only 1 is best for reliability in that you can always pick the best one, so 
if one is ruined by motion artifact you still have data for that subject.

cheers,
Bruce

On Sat, 8 
Mar 2008, Jeremy Gray wrote:

> thanks Peter, sounds very promising.
>
>
> On Mar 8, 2008, at 10:59 AM, Peter Kochunov wrote:
>
>> Jeremy
>> 
>> You might want to check out this paper.
>> http://www.ncbi.nlm.nih.gov/pubmed/16628607? 
>> ordinalpos=10&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.P 
>> ubmed_RVDocSum
>> 
>> We routinly average up 7-8 MPRAGE studies in order to reduce motion in the 
>> high-res anatomical images.
>> You can judge the quality of the individual segments by running a histogram 
>> analysis  on them
>> pk
>> 
>> ----- Original Message ----- From: "Jeremy R. Gray" <[log in to unmask]>
>> To: <[log in to unmask]>
>> Sent: Friday, March 07, 2008 11:13 PM
>> Subject: [FSL] combine two MPRAGEs?
>> 
>> 
>> Hi FSL'ers,
>> 
>> I'm new to FSL (really liking it so far). I'm hoping for advice on the best
>> approach to combining two MPRAGE images from the same subject. This is for
>> 100+ subjects, and so I will script it. the idea is to end up with one
>> higher-quality image to use in structural analyses. one image is from the
>> start of the scan session, and the other from the end (~1.25 hours later).
>> 
>> flirt seems like the way to go, so I searched the archives and the flirt
>> lecture notes from the web (pdf), but did not see something on combining
>> MPRAGES. my apologies if I missed it.
>> 
>> one question is:
>> - is it always better to combine two images? presumably there could be
>> pathological cases (e.g., lots of movement resulting in a blurry image)
>> where a single good MPRAGE is better than combining one good and one bad --
>> so is there a way to tell that you are in such a situation (especially for 
>> a
>> script to tell this)? just inspect afterwards?
>> 
>> using flirt seems straightforward:
>> - prior to flirt, run bet -B on each image (= my interpretation of flirt
>> lecture slide #45). but maybe for having the same sequence, the non-brain
>> stuff will actually help the alignment? and maybe doing bet on the combined
>> image will give a better extraction (for having a higher-quality input)?
>> 
>> - just pick one image to use as the reference, "better quality" should be
>> moot with 2 MPRAGEs (except in pathological cases)
>> 
>> - a rigid body 6-parameter model seems fine because the images are from the
>> same subject, same scanner, same day. is there any possible advantage to
>> more df for my situation?
>> 
>> - search option = "already virtually aligned" is probably fine
>> 
>> - cost function: correlation ratio is the default in the GUI. however, I've
>> heard that normalized mutual information is very good, in particular is
>> robust to small non-brain bits left over from brain extraction. any reason
>> not to use NMI (especially if I do bet prior to flirt)?
>> 
>> - trilinear interpolation (= default) -- any advantage to sinc?
>> 
>> thanks much,
>> 
>> --Jeremy
>> 
>> 
>> /*-------------------------------------------------------------
>>      Jeremy R. Gray, PhD
>>      Assistant Professor, Yale University
>>      Dept. of Psychology & Interdepartmental Neuroscience Program
>>      mail Box 208205, New Haven, CT 06520-8205 USA
>>      office SSS 212
>>      http://maps.google.com/maps?q=1+Prospect+St,New+Haven
>>      phone  203-432-9615 (office)
>>      fax    203-432-7172 (include Attn J. Gray)
>>      web http://www.yale.edu/scan/
>> -------------------------------------------------------------*/
>
>
>