[Insight-users] Re: Speed of B-Splines Deformable Registration

[Luis Ibanez]

Hi Andy,

The current implementation of BSpline registration
is not quite optimal.

There are a number of standard practices that we
are not taking advantage of in the current ITK
implementation.

For example, the computation of Metric Derivatives
could be done by considering only the area of influence
of a particular BSpline grid node, instead of computing
it over the entire image.

The group of Daniel Rueckert group at the Imperial College
of London is currently porting some of their optimized
BSpline registration to ITK classes, however, it may take
a while before they are fully integrated in the toolkit.

In the meantime, your best options are:

   1) Do multi-Resolution on the Image Data

   2) Do multi-Resolution on the BSpline grid
      (see example DeformableRegistration6.cxx)


Please let us know if you have further questions.


   Thanks


      Luis



----------------------
Andy Eow wrote:

> Hi,
>  
> I'm trying to run the DeformableRegistration4.cxx example which uses 
> B-Splines for non-rigid registration. My goal is to eventually replace 
> the means squares metric in this example with a MI metric. However 
> before that, I'm just simply running a slightly modified version of the 
> example first to take in a 3D volume (200x200x200) instead of a 2D 
> slice. After running for 2+ hours, this is all I get:
>  
>  F = 21.4132, GNORM = 1.62984
> *************************************************
>    I   NFN    FUNC        GNORM       STEPLENGTH
>    1    2        18.592       1.131       0.614
>    2    3        16.318       0.835       1.500
> With a maximum iteration number at 1000, this process is going to take 
> quite a few days to complete. I'm running this on a Pentium 1.6M with 
> 1GB RAM. Just wanted to do a sanity check to see if this sounds normal. 
> If not, what kind of worst case running times should I be expecting on a 
> 3D volume of size 200x200x200?
>  
> Cheers,
> Andy
>