Application I : Deformable Models

Among the different techniques for segmenting medical images, the methods based on deformable models have the advantage of providing a strong a priori knowledge about the structures to segment. Also these methods usually require some level of user supervision to guide the segmentation. In many cases, this medical supervision is not a drawback (and even very beneficial) as long as the computation of the deformable models is fast enough to perform the segmentation task in a few minutes.

In Figure 2 we present some snapshots of image segmentation on 2D, 3D and 4D medical images. For the 2D image segmentation, the user can drag the contour with the mouse and propagate each deformable contour to neighboring slices. Each active contour can change its topology when collisions are detected (). The 3D deformable surfaces are represented as simplex meshes () which are deformed inside volumetric images. The slice view of deformable surfaces can be seen in any orthogonal slicer and user interaction (dragging the surface towards or away an image region) can be performed in the 3D view or in each slice view. We found that the combination of 3D and slice views is very helpful for performing the difficult 3D segmentation of structures such as the kidney or liver. Indeed, the 3D view helps the user to understand globally the segmentation process whereas each orthogonal view allows to control the segmentation accuracy in specific planes. Deformable model based segmentation benefits a lot from the use of scripting language because we can easily set and modify the large number of parameters for a given application. For instance, we have created a set of scripted applications for segmenting the kidney with some user guidance such as definition of the ROI, image windowing, selection of a kidney template. However, the computation complexity is hidden by the interface so that the user do not have to enter any magic parameters.

The 4D (3D+T) deformable model interface is based on the interface for a 3D deformable model with the addition of a slider that select the time at which the scene (consisting of a volumetric image and a deformable surface) is viewed. Also a 4D image viewer (see Figure 7) can be synchronized with this 4D camera and thus allows to see the trace of each surface model in the image.

Figure 2: Left: surface simplex mesh duality with triangulations. Right: simplex mesh geometry.

[2D DeformableModel]     [3D Deformable Model] [4D Deformable Model]