Dear Nuria,

> Dear FSL experts,
> 
> While trying to optimise our DTI preprocessing pipeline, I have been testing the results of eddy. I have run eddy with the same exact options on the same data various times, and the result I obtain is different at each trial. Can that be due to the program starting with different initialisation parameters at each run? What else could be the reason?
> 
> I have previously run TOPUP on my data, and the results of TOPUP are always the same. My DTI study has one B0 image and 32 images with b=1500. The 32 directions are all on the same half-sphere. eddy is run with the default options.
> 
> All the output files of eddy change to some degree at each execution. Even the outlier report has different values. The results tend to be close but different. For instance, the slices with outliers are almost the same, and the number of std deviations is usually similar but not equal, often changing at the second or third significant figure. At the rotated bvecs file, even if most values are close, that is not always the case. And in the corrected image there are considerable differences in intensity (>10% in some brain voxels, often >5%).
> 
> Do you know if there is any explanation to this?

there is a random component in eddy. The subset of intracerebral voxels that are used to estimate the hyperparameters for the Gaussian process are drawn at random, and different runs of eddy will produce different random sequences of voxels. It should have very little impact on the estimated movement and eddy currents, and hence on interpolated values etc. In your case you have relatively few directions (32) and are on the half sphere, so potentially the effects will be a little greater than I would normally see.

One way to make the impact smaller is to increase the number of voxels used. You do this by setting the parameter --nvoxhp to a larger number (the default is 1000). There is also a parameter --initrand that means that the random number generator will always be seeded with the same number, so would guarantee the same result each time.

Just remember, just because the results is the same each time doesn’t make it more correct. I personally find it useful to be able to see the impact of the hyperparameter estimation on the end results.

Jesper

> 
> Thank you very much in advance!
> 
> Best,
> 
> Núria Roé
> 
> 
> Unidad de Imagen Molecular, CIMES
> Fundación General de la Universidad de Málaga 
> C/ Marqués de Beccaría 3, 29010 Málaga, t 952 137 406/407  FAX 952 137 416
> 
>