Hi Brian
I had a look at your data, the problem is the units of the b values that you have used. We assume units of s/mm2. In your case these seem to be six orders of magnitude off, so bedpostx assumes that they are very small. I can see that the units are not clearly specified in the fslwiki, I will have that changed.
What happens is that FDT then attempts to drive the diffusivities to high values to compensate for that. Dtifit does that OK (you can check the MD values), but bedpostx has priors on the parameters and these compete against the high diffusivity values. Therefore, results re highly unstable. The differences that you see between the two datasets should be due to the different burn-in periods that you have employed (twice as much in one of the two runs), causing the (anyway unconverged chains) to sample different areas of the parameter space.
To summarise, if you multiply your b values by 10^6 and rerun, the problem should be resolved.
Cheers
Stam
On 12 Feb 2013, at 22:58, Brian Stirling wrote:
> Hi Stam,
>
> I do not encounter the same issue with DTIfit. This is definitely strange.
>
> Thanks,
>
> Brian
>
|