 |
 |
Dear Daniel, It's been over a decade since I looked into the motion correction code. With hindsight, I would guess that the optimal amount of smoothing should depend on the separation between sampled voxels and the amount of noise in the images. The separation should also be as small as possible - although this is likely to slow things down. A slightly different model should be used for dealing with larger, or anisotropic, voxels - similar to the types of approaches used for resolution recovery in baby brain imaging. Again, this would slow things down even further. Then there are other aspects to deal with, such as distortions (EPI as well as motion of the brain-stem) and the noise not being i.i.d Gaussian. Also, some of the noise in the motion correction model is the actual signal of interest in other steps in the analysis pipeline. One day, I might have the chance to revisit at motion correction and make it better, but this would largely be a thankless task. Proper exploration of motion or other confounding effects (explanations of the data that are not related to the "question") is often not regarded as proper neuroscience (see the bit about rats and mazes in http://calteches.library.caltech.edu/51/2/CargoCult.htm ). Best regards, -John On 22 September 2015 at 17:49, Daniel Gallichan <[log in to unmask]> wrote: > Dear John > > I’m a post-doctoral MR physicist at the EPFL in Lausanne working on > motion-correction, and I posted this question recently to the SPM > discussion forum but I haven’t received any replies. As you are listed as > the author of ‘spm_realign’ in the Matlab code I was wondering if you might > be able to help - or to point me to someone (or even a reference) so I can > try to understand these parameters better? > > Thanks very much > > Daniel > > > > > — > > Hello to the SPM community > > I have recently been using SPM's 'realign' tool to estimate motion parameters for image-navigators (in my case, FatNavs acquired using a fat-excitation) which are then used to retrospectively correct the 3D k-space of the structural host sequence to reduce motion artifacts in high resolution imaging (some more details available here: http://www.cibm.ch/page-117748-en.html ). > > In attempting to find the 'optimal' parameters for my navigator (in terms of voxel size and parallel acceleration) I have discovered that the estimated motion parameters are quite dependent on the choice of the 'sep' and 'fwhm' parameters in realign. Would someone please be able to describe how I should choose the values for these parameters? - The SPM manual implies that 'sep' should be as small as possible - but for FWHM suggests only to use 7mm for PET and 5mm for MRI, and doesn't put this into perspective with respect to voxel size. I assume here 'MRI' refers to fMRI experiments with large EPI voxels... > > The specific experiment which I have completed to test various navigator parameters involved acquiring data with several subjects making small head movements while 2mm images were acquired. I then downsampled these images to various resolutions and acceleration factors, and would like to see how far this can be done while still getting good motion estimation. Should I use the same 'sep' and 'fwhm' for all image resolutions? > > A further consideration is that of the 'register to mean' option. This would seem to be a good idea in order to get the most robust motion estimates - but in this validation experiment comparing different navigator resolutions I suspect it may introduce a bias in the comparison if the 'zero-motion' case is not quite the same for each image resolution. Am I right to think this could be a problem? > > Thanks for any help and suggestions you may have for this. > > > Regards > > > Daniel > > > — > > > > -- > Daniel Gallichan, DPhil > Post-Doctoral Researcher > Centre d'Imagerie BioMédicale (CIBM) > EPFL - SB IPSB LIFMET > Station 6 > CH-1015 Lausanne > Switzerland > > Tel: +41 21 693 7651 > >