[Insight-users] Re: [vtkusers] A theoretical question

[Daniel Mace]

Arne,

If you want, you could try the filter and see what the shape models look 
like.  One of the benefits of levels sets are that they are capable of 
handling topological changes rather transparently.  If you have multiple 
disconnected regions, the PCA step will still build you a model with 
these multiple regions (I haven't tried personally, but I don't see why 
it wouldn't).  However, depending on your data, the ability to pre-align 
objects, etc., it may be that it doesn't effectively capture your 
shapes.  This is one of the areas that coupled shape models try to 
overcome, by creating separate level set models and providing some kind 
of covariance model linking them.

This isn't to discourage you-- most models do make some underlying 
assumptions/simplifications.  The ideal algorithm would separately align 
each individual segment and then build some model that captures the 
covariance between shapes, distances, poses, texture qualities, etc.  
Naturally, this isn't a simple problem to model, and even if it was, the 
amount of data needed to accurately model something of this nature would 
probably be quite extensive.  Tsai et al. is a great extension to 
Leventon's work, but it too does make some simplifications that might 
not capture all of your shape variations.  It would probably be more 
suited in your case though, and if you really would want to use it, you 
would probably need to implement it (or email the authors and see if 
they will send you the software for it).

Cheers,
Dan


Arne Hansen wrote:
> Dear Daniel
> Thank you for your answer. This looks very interresting I think.
> Could this filter also be used for a coupled shape model i wonder?
>  
> A coupled shape model is where i have several regions represented in 
> my same shape model. This means that i simulataneously can segment 
> many regions, and use the prior knowledge of the spatial position of 
> the different regions, which i have also captured in my shape model.
> For more information of the strongly coupled model, this article gives 
> a good insight:
> http://ssg.mit.edu/group/alumni/atsai/Papers/IPMI03.pdf
>  
> I hope to hear from you or anybody else soon.
> Yours Sincerely
> Arne Hansen
>
>  
> 2007/1/26, Daniel Mace <dlm19 at duke.edu <mailto:dlm19 at duke.edu>>:
>
>     Arne,
>
>     Is your question: what methods are out there that you can use (but
>     would
>     need to implement), or what methods are in ITK/VTK that you could
>     use?
>     The standard shape model in ITK is the the work done by Leventon
>     et al.
>     (
>     http://spl.bwh.harvard.edu:8000/pages/papers/leventon/cvpr00/cvpr00.pdf
>     <http://spl.bwh.harvard.edu:8000/pages/papers/leventon/cvpr00/cvpr00.pdf>
>     ).  His method basically relies on the observation that the
>     gradient of
>     an image is inversely related to a signed distance map and created
>     a MAP
>     estimate from it.  If you already have your mean + modes of variation
>     then using his method should be easy to implement (easy to implement,
>     getting the parameters to converge might take a bit).  It's
>     described in
>     9.3.7 in the ITK software guide.
>
>     Cheers,
>     Dan
>
>     Arne Hansen wrote:
>     > Hello. I have a question regarding the use of shape models built
>     using
>     > Signed Distance Maps(Level Set Function).
>     > I have applied a principal component analysis on my Level Set
>     > Functions, and found a meanshape + modes of variation.
>     > I now need to use these as prior information in a segmentation model
>     > where certain regions(those which my shape model represent) needs to
>     > be segmeneted.
>     > I am however unsure of how i can do this. So my question is if an
>     > Active Appearance Model technique can be used when i have built my
>     > shape model on Level Set Functions, or if i actually need to do
>     it on
>     > landmarks.
>     >
>     > Is there other methods to use?
>     > Maybe you can recommend a method to use.
>     > Thank you very much, and best of regards.
>     > Arne Hansen
>     >
>     ------------------------------------------------------------------------
>     >
>     > _______________________________________________
>     > This is the private VTK discussion list.
>     > Please keep messages on-topic. Check the FAQ at:
>     http://www.vtk.org/Wiki/VTK_FAQ
>     > Follow this link to subscribe/unsubscribe:
>     > http://www.vtk.org/mailman/listinfo/vtkusers
>     >
>
>