Hi -  
As Matt said, It is perfectly reasonable to expect more fibres for higher angular resolution data, since the "allowed" angular separation between two fibres gets smaller, and hence you can start modelling fibres bending within a voxel etc. (Andreas: that might be what is happening in the corpus callosum). 

There are two main issues related to the ARD:

1. With more data, it might be necessary to increase the length of the burnin period. 
For people interested in the technical details: this is because bedpostx uses a Metropolis Hastings MCMC sampling. Basically, it samples by taking steps in parameter space. If you have more data, you need to take smaller steps (as more data gives "large" jumps in the likelihood). So it may take longer to "kill" unnecessary parameters by sampling near zero.
You can test this by running longer burnins in a small mask where you expected one fibre but got two (let us know what you get).

2. Having more data emphasises modelling errors (e.g. due to subject motion, rician noise, multiple diffusivities, etc.) - 
This is why we allow the user to up-weight the model (or down-weight the data) using the ARD weight. This parameter was tested for standard data where it works well, but might need some tuning in your case.


Cheers,
Saad.




On 7 Apr 2010, at 11:32, Jeurissen Ben wrote:

> Part of it are "averages", but with quite different bvec rotations. Any
> idea why this is a problem for the ARD?
> 
> 
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
> Behalf Of Matt Glasser
> Sent: dinsdag 6 april 2010 20:59
> To: [log in to unmask]
> Subject: Re: [FSL] bedpostx with high angular resolution dataset
> 
> The length of time is certainly because of the large number of
> directions
> (even 256 takes a long time).  Do you have several averages with
> slightly
> different bvecs rotations after eddy_correct or something like that?  I
> think I remember Saad saying that ARD doesn't do as well with this (it
> is
> better to average to improve SNR).  With large number of directions it
> is
> possible to reconstruct at least 3 fibers in many voxels (this will
> likely
> be more accurate).  
> 
> I think it is expected to find a lot more multi-fiber voxels in datasets
> with large numbers of DW directions.  Very high angular resolution data
> should be more sensitive to lower strength subsidiary fibers & larger
> numbers of subsidiary fibers.  Also, One thing that can happen is that
> broader uODFs can be split into multiple fibers (potentially
> incorrectly).
> The good news is that this is not necessarily bad for tractography!
> Because
> this is an area of active work, I hope Saad or Tim will correct anything
> I
> have said that is incorrect. :)
> 
> Peace,
> 
> Matt.
> 
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
> Behalf
> Of Ben Jeurissen
> Sent: Tuesday, April 06, 2010 1:27 PM
> To: [log in to unmask]
> Subject: [FSL] bedpostx with high angular resolution dataset
> 
> Hi,
> 
> I'm running bedpostx on a DW dataset with 700 DW directions (after
> b-matrix
> rotations), b-value 1000. Apart from a very long processing time (which
> is
> probably due to the large amount of DW directions), everything runs
> fine.
> 
> However, the automatic relevance determination doens't seem to do its
> job
> well. For example: bedpostx finds 2-fiber voxels almost everywhere in
> the
> dataset. It even reports quite some multi-fiber voxels deep in the
> corpus
> callosum where I would not expect them and the neighborhood  does not
> support them at all.
> 
> Does this mean I have to reduce the ARD weight? If so, what value do you
> suggest I set it to? Or is the method not suitable for datasets with
> large
> numbers of images?
> 
> Kind Regards,
> Ben
> 

--
Saad Jbabdi
University of Oxford, FMRIB Centre

JR Hospital, Headington, OX3 9DU, UK
(+44)1865-222466  (fax 717)
www.fmrib.ox.ac.uk/~saad