Dear Ben,
I have had a look at your data and I could replicate the error. It is now clear to me what happens, but unfortunately I don’t have an immediate solution.
It has to do with the fundamental (simplified) model we have of how an off-resonance field affects EPI images. When there is a strong gradient in the field (which essentially means that neighbouring voxels are displaced by “very” different amounts) it leads to pile-up of intensity. This pile-up is modeled using the Jacobian of the displacement map. If for example the Jacobian is 0.1, which means that a voxel has been squashed to a tenth of is original size we expect the intensity to have been modulated by factor ot ten. But what happens when a voxel has been squashed to zero size? According to our simplified model that means that the intensity is now infinite, which is clearly nonsense.
Hence, our simplified model breaks down when the jacobian approaches zero. When estimating the field using topup this will never happen because the same approximate model is used in topup as in eddy. But when measuring a field using a traditional fieldmap this can happen. This is the reason that it is generally advised against using Jacobian modulation in for example fugue (the --icorr flag).
In your fieldmap there are a couple of voxels at the front edge of the pons/optic chiasm ares in slices 22 and 23 where this happens. It drives the predictions in eddy to unreasonable values and causes the whole process to diverge from the first iteration.
This was not something I had foreseen and I need to have a good think about how best to prevent this from happening.
Jesper
On 23 Oct 2016, at 22:13, Chernoff, Ben <[log in to unmask]> wrote:
> Hi Jesper,
>
> Were you able to replicate my error with the fieldmap that I uploaded on Thursday?
>
> Thanks,
>
> -Ben
>
>> On Oct 20, 2016, at 1:28 PM, Ben Chernoff <[log in to unmask]> wrote:
>>
>> Hi Jesper,
>>
>> I just uploaded the fieldmap (converted to Hz, smoothed with Fugue, flirted into diffusion space) that I fed Eddy when receiving the error, as well as the nodif brain mask.
>>
>> Thanks again for looking into this!
>>
>> -Ben
|