Hi Darren and others
Sorry to join the conversation late - being more involved in MR physics I have to confess I don't actually subscribe to the FSL list...
I have just a few comments to add.
Firstly, I definitely agree with Darren that you are seeing the same vibration artefact as we observed on the system in Oxford prior to the Siemens hardware adjustment. However, if you see an effect with |x| as low as 0.45 then I imagine that you have an even bigger effect than we observed - which could be due to small differences in hardware, or even in how you restrain the head of the subject (we didn't test it thoroughly, but it certainly appeared to be the case that by holding the subject in place on the sides by applying pressure to the headphones it also creates a very good mechanical connection for transmitting the vibrations...). Unfortunately, a bigger effect also means you are less likely to be able to get a reasonable correction.
The correction method described in our paper is unfortunately also only an approximation. The Tukey windowing function then approximates the effect of the k-space filter which is automatically applied to the 'smaller half' of the 3/4 Fourier data. I tried applying the technique to a couple of 30-direction datasets and observed a similarly incomplete correction to that which you show in this thread. Having more directions effectively means that you are able to sample more of this curve arising from the filter, and are likely to get better results. It is also worth considering that even if the FA maps look much better after the correction, the resulting diffusion tensor in these regions will be less reliable than if the artefact were not present - as the extra parameter will necessarily increase the confidence interval on all estimated parameters. It is certainly preferable to not have the artefact in the first place...
If you don't need to do direct comparisons of tensor values within the affected regions, then the simplest and most reliable method to deal with the artefact would be to identify affected regions by their residuals from a normal tensor fit, and then exclude these regions from the datasets. Subjects where the affect region overlaps with the region of interest would sadly then need to be discarded - but that may be the only option left...
You mentioned doing a voxel-by-voxel fitting approach - do you mean effectively having different Tukey filter parameters for each voxel? If so, I don't know how you would choose which values to use. Ideally you need to know how the signal varies with |x| when there is no diffusion present (or isotropic diffusion) which without vibration should be constant. With vibration, however, this will become location-dependent. The method we presented makes a guess at this function by averaging over all the affected voxels which have been identified. I don't see how you could do this on a voxel-by-voxel basis.
My general advice for avoiding the artefact is, as Arman mentioned, using a full-Fourier acquisition with the necessarily longer TE. When we tried this with iPAT=2 the TE was only marginally increased, and the SNR appeared to be quite similar (getting a precise SNR comparision is pretty tricky - especially where parallel imaging is involved...).
Hope this is of some use - and I hope you can still get something useful from your data!
Cheers
Dan
|