 |
 |
Hi Dianne, I mostly agree with Matt's answer. I would add that, in my experience, TE has the most important effect in the SNR, so everything that you can do to reduce the TE is going to help a lot. The largest contribution to the TE comes from the diffusion-encoding gradients, so stronger gradients are a plus, as well as trying to collect data with the single spin-echo sequence (although, as Matt points out, the eddy-current-induced distortions will be worse). Other factors, like reducing bandwidth, parallel imaging, etc. will have less effect in the TE, although they will help (and you also get less EPI distortions). I am totally against partial-fourier imaging for diffusion, since you are assuming that the phase of the object is zero and doesn't change during the readout, so the two halves of k-space are complex conjugates of each other. In my experience, the phase of the object does weird things, probably due to mechanical vibrations. Evidently, if you collect more data (6/8 or 7/8 partial fourier) you get more information to retrieve the correct phase, but you can get bad artifacts (like the "holes" in the brain that people have reported with Siemens scanners: I never use partial fourier and never got the artifact). I would also add that, for the same number of volumes, collecting more directions is better (less bias estimates) than collecting less directions but with more averages, as Derek Jones has shown (I can find out the reference, if you want). Other than that, Matt's answer is quite thorough. Good luck, -Pablo On Wed, 21 Apr 2010 18:34:25 -0500, Matt Glasser <[log in to unmask]> wrote: >1) Subject motion (even after eddy current correction) is certainly >highly correlated with poorer bedpostx reconstructions (fewer subsidiary >fibers, more uncertainty in fiber orientations), so it would make sense that >it would negatively influence FA estimates. > >2) If the problem is severe you probably should exclude the subject. > >3) SNR scales linearly with magnet strength, so you would expect lower >SNR for a given number of DWIs acquired at 1.5T vs 3T, all other things >being equal. > >4) This is a more difficult question: parallel imaging itself tends to >decrease SNR a bit; however, it also allows you to decrease the TE, which >increases the SNR and reduces the distortions. The same is true of partial >fourier. > >5) One tends to get worse SNR with higher bvalues, but better contrast >to noise ratios (CNR). For FA, I think it has been established that around >b=1000 is optimal (SNR vs CNR), but much higher bvalues are optimal for >estimating crossing fibers. The exact optimum bvalue will also depend on >your hardware, sequence, and imaging resolution, because all of these affect >the SNR term. > >6) Averaging multiple scans will increase SNR and reduce noise, so >long as you properly corregister them and remove subject motion within the >timeseries. > >7) SNR varies with the number of DWIs acquired, so as long as you make >up for fewer directions with more averages, the SNR should be the same. >That being said, your estimates of fiber orientations (and especially >crossing fibers) will be worse with fewer directions, so it is always better >to acquire more directions (unless you want two averages for something like >phase up/phase down distortion correction). > >8) I'm probably the least clear on what this does to SNR. It is >really bad for tractography because it allows for the creation of >intermediate orientation fibers that can cause inaccuracies, but I don't >know how it affects the accuracy of FA values. > > > >Other issues you have not considered: > > > >9) SNR will vary depending on the coil you used (more channels are >better), and the gradient set in the magnet (stronger gradients allow a >lower TE and thus higher SNR). > >10) SNR will vary depending on if you use a dual spin echo sequence vs a >single spin echo, again because you can use a lower TE with a single spin >echo (but eddy currents will be worse). > >11) Imaging resolution has a big impact on SNR, of course. > >12) Higher imaging bandwidth generally results in worse SNR but less >distortions. > > > >Peace, > > >Matt.