I only have data from the human connectome project, I usually don't collect data myself. Right now I'm only interested in the motion correction, not the eddy current correction, as I just recently started with DTI (worked with fMRI for a long time).
-> Ok, that data comes with different levels of processing, though, including something that is supposed to be fully processed? I am not an expert in it, but there are numerous special factors to consider when processing, such as the fact that the data was acquired at 7T and therefore contains very nonlinear distortions (different than those seen at 3T), etc. Before you do a comparison with methods, you might want to contact the authors of the various preprocessing tools to get their advice.
NB: it is also probably not really possible to think of each distortion on its own, as they all feed into each other. That makes processing extra tricky, but I think it is probably the reality of the situation.
I have a pretty good registration algorithm, and I want to test how it performs compared to other algorithms for DTI. For DTI data, there are large differences in the image intensity and the noise level for different b-values. This makes it hard(er) to correct for motion between different b-values, like 1000, 2000 and 3000. My registration algorithm is designed for handling large intensity differences, and a quick test shows that it is more robust than flirt in FSL for my DTI data (but I don't know how common it is to use flirt for DTI).
-> Great, always good to have good registration. Yes, DTI is different than FMRI because: the contrast is so different between DWIs and b=0 volumes, and each DW volume has very different contrast to every other one (with a different gradient).
The FSL tool 'eddy_correct' is really just a linear affine registration tool that is a wrapper for flirt, registering each DWI to the b=0 volume, using the default cost function. I think the Rohde et al. (2003) paper showed that the cost function of registration is important, and they used one optimized to do no warping when there is no motion (and I think this is what is implemented within TORTOISE). The flirt default cost function is (currently) not optimized for this, and so it can lead to stretching/warping even in the presence of no motion, merely due to the differences in b=0 and DWI contrasts. So, you should consider the cost function carefully for registration.
--pt
Edited 1 time(s). Last edit at 09/19/2015 05:26PM by ptaylor.