[Insight-users] registration

[Luis Ibanez]

Hi Imho,

I'm affraid that what you are looking for, is not
available in the toolkit at this point.

The Model to Image registration approach is not
a Point based registration. It is not associating
points from two point sets as ICP does.

Instead you have a geometrical model and you
define your own metric that will measure how well
the model match to an image.

PointSets are one among many other possible
representations of SpatialObjects.

You may want to look at the Model Based Registration
section of the SoftwareGuide

Section 7.14, pdf-pages 234-244.


This algorithm is fitting a geometrical model to
an image.


The problem with ICP is that there is a lot of
time spent in finding point correspondances.
Actually most of the time goes wasted in this
stage of the algorithm. This time grows to the
square of the number of points unless you use
some kind of auxiliary data structure like a
PointLocator.

If you imagine to build an image with a distance
map of one of the point sets, and then registering
the other point set against this image, you will
visualize better why Model to Image registration
may be more efficient than Model To Model registration.

Note that the group developing this techniques is
actually doing Liver registration for image guided
intervention.  In this context, what you want to do
is to create a geometrical model of the Liver, using
the SpatialObjects available in:

             Insight/Code/SpatialObject

Then register such model against an image.

Modeling is probably the next step in the evolution
of medical image algorithms since it allows to
introduce anatomical meaning to the data representation.

Note that PointSets and Image are not aware of
representing a Liver, a Hearth or a Lung. SpatialObjects
on the other hand can be built with growing complexity
using a CSG-kind of grouping, making possible to generate
meaninful shapes.



Please let us know if you have further questions.



Thanks


   Luis



----------------------------------------

imho wrote:

> Hi Luis,
> 
> you said that Iterative Closest Point wasn't implemented, so wich 
> algorithm is it? Iterative Inverse Perspective? Another one?
> 
> thanks
> imho
> 
> Luis Ibanez wrote:
> 
>>
>> Hi Imho,
>>
>> At this point probably the more interesting
>> method is Model to Image registration.
>>
>> This is done right now in ITK by using the
>> SpatialObject classes for representing
>> geometrical models. An example on Model
>> to Image registration is available in the
>> SoftwareGuide.
>>
>>
>> Does this help to answer your question ?
>>
>>
>>    Luis
>>