[Insight-users] Demons Registration

[Ganesh Narayanasamy]

As a follow-up to my earlier message, I can notice that the MSE & Harmonic energy tapers off after just 20 to 30 iterations. This is pretty fast compared to what I have seen with other methods and implementations. However, I have not checked with all permutation and combination of parameters that are possible here.
I suppose these are the only 2 metrics that we need to watch, apart from making sure that the Jacobian stays positive. I am using the same parameters given earlier with Deformation field sigma = 1.5 (default)
I can also see that adding to the multi-resolution levels seems to help, mainly in getting a higher harmonic energy. 

Apart from viewing the resulting image and the transformation, I would like to measure the registration error which helps in 
understanding how "good" is the warping. One 
method is to pick a few fiducial point pairs in the 2 original images and apply the transformation field onto the fiducials of the fixed image.  Is there a way to do so here? I checked with the example code and could not find any.

Appreciate your comments... thanks,
Ganesh

> 
> > From: tom.vercauteren at m4x.org
> > Date: Thu, 24 Feb 2011 08:47:01 +0100
> > Subject: Re: [Insight-users] Demons Registration
> > To: nganesh76 at hotmail.com
> > CC: insight-users at itk.org
> > 
> > Hi Ganesh,
> > 
> > On Wed, Feb 23, 2011 at 19:04, Ganesh Narayanasamy
> > <nganesh76 at hotmail.com> wrote:
> > >
> > > When I run DemonsRegistration using the options as given in the command line
> > > below, I can get a few output files like final image, deformation field etc.
> > > How would I get the rest of the output like "True field file"?
> > 
> >  The "true field" is an input not an output. It may be used for
> > synthetic experiments where the actual deformation is known.
> > DemonsRegistration then compares at each iteration its result against
> > this ground truth (but of course it does not use it to drive the
> > registration).
> > 
> > 
> > > As the image was not deforming enough, I had opted to reduce sigma from the
> > > default value of 1.5 to 0.8.. is this the correct approach?
> > 
> > This entirely depends on your use case... That being said, I would
> > first try to use more iterations and then reduce the sigma.
> > 
> > 
> > > Size of the fixed and moving images are 384x384x64 pixels each. I would like
> > > to register using the multi-resolution approach starting with 1/4th number
> > > of pixels in all 3 axes (i.e, 96x96x16), going to 1/2 (ie., 192x192x32
> > > pixels) and finally to the full image. How will I do this?
> > 
> > It is already done, see the number of multiresolution levels and
> > number of iterations per level:
> > >   Number of multiresolution levels: 3
> > >   Number of demons iterations: [ 15 10 5 ]
> > 
> > Hope this helps,
> > Tom
> > 
> > 
> > > Command:
> > >>DemonsRegistration.exe -f CBCT1.mha -m CBCT4.mha -o Final.mha -O DefFld.mha
> > >> -s 0.8 -d 1
> > >
> > > Starting demons registration with the following arguments:
> > > Arguments structure:
> > >   Fixed image file: CBCT1.mha
> > >   Moving image file: CBCT4.mha
> > >   Input field file:
> > >   Input transform file:
> > >   Output image file: Final.mha
> > >   Output field file: DefFld.mha
> > >   True field file:
> > >   Number of multiresolution levels: 3
> > >   Number of demons iterations: [ 15 10 5 ]
> > >   Deformation field sigma: 0.8
> > >   Update field sigma: 0
> > >   Maximum update step length: 2
> > >   Update rule: exponentialize and compose (Diffeomorphic)
> > >   Type of gradient: symmetrized (ESM for diffeomorphic and compositive)
> > >   Use histogram matching: false
> > >   Algorithm verbosity (debug level): 1
> > >
> > > Thanks,
> > > Ganesh
>  		 	   		  

 		 	   		  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20110228/1d32fa8a/attachment.htm>