[Insight-users] How do I validate a 3D segmentation process ?

[Luis Ibanez]

Hi Mathieu,

Validation is an ill-defined problem.

You can probably rephrase it in terms
closer to Popper's view of scientific
research, and just say that you plan
to contrast your methods against reality
in order to increase your level of
confidence on them.

In this context, the two traditional
mechanisms for validation (read as
"Increase level of confidence") are:

1) Create syntetic images with know
    size shapes. Segment them with
    method X and contrast the values
    to those used when generating the
    image.

2) Recreuit human operators with some
    background in anatomy. Train them
    in manual segmentation techniques
    and convince them (via persuasion,
    pizzas or even salary) to pass hours
    in front of the screen segmenting
    a set of defined volumes.  Ask them
    to repeat the segmentation several
    times in order to evaluate their
    variability.

    Then you compare the segmentations
    done with your method X against their
    manual segmentations.


Visualization alone is unfortunately
not a good mechanism for validation.


Registration between meshes an images
are possible. At this point the best
mechanism for doing so is the Model
based registration implemented in
StatialObjectToImageRegistrationMethod.

This topic is discussed in the
SoftwareGuide.



Luis




---------------------------
Mathieu Malaterre wrote:
> Hi all,
>     I have a general question about validation in segmentation, and I 
> would like you, people, to comment on this.
>     I have done filters that segmented my images. I am then using a 
> software to see (at the same time) the model and image. And everything 
> 'looks' ok. But now if I play little with parameters I can find other 
> results that 'looks' good.
> 
>     So my question is what is the method to validate a 'good' 
> segmentation (just looking at images is too biased).
> 
>     - As I am doing 3D segmentation I don't want to do any 2D related 
> stuff.
>     - Has anyone tried registration between Mesh and Image ?
>     - I have thought about using Mesh (mesh.epfl.ch) but I can't mesure 
> error to an isovalue image (I am using a dynamic isovalue to find my 
> model).
> 
> Thanks,
> mathieu
> 
>