Hi
> 1) How is "evidence" for more complex fiber structure measured in the automated relevance determination.
Bedpostx uses a 1/x prior on the volume fractions (by default only on the fibres 2,3,etc.) This prior competes with the likelihood function (data fit) and forces the volume fraction to near zero if that doesn’t come with too much penalty in the fitting.
More details in Behrens et al. NeuroImage 2007.
> 2) If a crossing fiber exists in a voxel but has degenerated partially such that the "evidence" that this more complex model exists decreases, would residual diffusion data of this lost fiber direction contribute to the uncertainty of the primary fiber direction?
Indeed, if there is variance to be explained by an additional fibre, and if the a.r.d. kills that additional fibre, you will get higher uncertainty on the primary direction. If the a.r.d. prior works this should not happen though.
Cheers
Saad
>
>> I have read that the ARD estimates these parameters independently, such that one parameter's prior and posterior distributions do not affect the other parameters' distributions. I believe this could answer my second question. However I am not absolutely convinced and understanding conceptually an answer to #1 would help tremendously.
>
> Thank you!
>
> ----
> Miguel
>
> ________________________________________
> From: FSL - FMRIB's Software Library <[log in to unmask]> on behalf of Miguel Sotelo <[log in to unmask]>
> Sent: Wednesday, September 07, 2016 10:53 AM
> To: [log in to unmask]
> Subject: [FSL] Interpreting dyads_dispersion (loss of fibers?)
>
> Hi all
>
> I am comparing a pathological group to a neurologically intact group, using some bedpostx outputs.
>
> I understand the bedpostX algorithm uses automatic relevance determination (ARD) to fit the more complex model (say, 3 crossing fibers) and determines whether there is *evidence in the diffusion data to support this complex model. If there is no evidence, then it reduces the model (say, to 1 fiber). In addition, I can see that the dyads_${i} images show the theta&phi distributions, and the dyads_${i}_dispersion shows the dispersion of the specific fiber-crossing orientation.
>
> My question is two-fold: 1) how is the "evidence" measured in the ARD approach, (when determining whether a diffusion data supports a complex model)? Also, how is "dispersion" measured in the dyads images? 2) more specifically, say anatomically there are two fiber populations in a voxel. After a pathology, one of those fiber directions is lost. As a result, you may expect the ARD to estimate a 1-fiber population at that voxel (which would be correct). However, would the "residual" diffusion of the now-lost second fiber population contribute to the dispersion of the still-intact first fiber population?
> *In other words, how independent are the fiber orientations estimated, and, if 2nd or 3rd fiber populations are not supported, do their components get "leaked" into the uncertainty of the first fiber population?
>
> As always, I really appreciate the time and the insight.
>
> Respectfully,
> Miguel
