Deformable models for image analysis

( J.-O. Lachaud , Annick Montanvert )

We have proposed a new highly deformable surface model to recover shapes from 3D biomedical images. Its geometry is explicit and defined by a triangulated mesh with distance constraints (as opposed to implicit level-set techniques). Simple distance criteria and local mesh reconfigurations ensure the model keeps a correct topology wrt its geometry. Experiments on various kinds of images have shown the interest of the approach.

Volumetric image segmentation and shape recovery using a topologically adaptable triangulated surface.

During my PhD, my objective was to recover shape from volumetric images. I have especially investigated methods that required no a priori knowledge on the final result (with regard to the topology for instance). Therefore I have designed and developed a deformable model based on a triangulated surface and constrained by internal and external forces, which can automatically adapt its topology to its geometry with very simple distance criteria. The model has proven to be robust on several kinds of images, both medical (CT, MRI, MR angiography) and biological (cell nuclei obtained by confocal microscopy). Moreover a coarse-to-fine approach speeds up the convergence process.

Related publications

[ Lac99a ], [ Lac98b ], [ Lac96a ], [ Lac96c ], [ Lac96d ], [ Lac95 ], [ Lac94 ]

MPEG animations of deformable models

cube animation (2.2 Mo). The volumetric image is a potential field representing a cube in space. The model is looking for iso-surface 0.5 in the field. Curvature constraints are progressively increased.
chain animation (0.6 Mo). The volumetric image is a potential field representing two intertwined torii. The triangulated surface is initialized as a big bubble around the torri. Its shape, at the beginning homeomorph to a sphere, is homeomorph to two torii at the end of the deformation process.
chain-grid animation (0.6 Mo). The volumetric image is a potential field representing two intertwined torii. The triangulated surface is initialized as a set of 1000 small bubbles scattered in the image. The deformation process transforms the triangulated surface into two intertwined torii.
skel-dir-1 animation (1.0 Mo). The volumetric image is a human head CTscan. The skull can be defined as an iso-surface within this image. Internal constraints are added to smooth the surface. The triangulated surface is initialized as a bubble embracing the image (with about 140000 vertices).
skel-pyr-1 animation (1.0 Mo).
skel-pyr-2 animation (1.0 Mo).
Same image as above. The triangulated surface is initialized as a bubble embracing the image (with only 3000 vertices). It first evolves in a coarse approximation of the image. A sketch of the skull is thus quickly extracted. Then the surface is globally refined and it evolves in a finer representation of the image. After this coarse-to-fine approach to the shape recovery problem, the surface has about 130000 vertices.
angio-pyr animation (2.5 Mo). The volumetric image is a phase contrast MR angiographic image highlighting the brain vessels. The surface is initialized as a bubble embracing the image (with about 3000 vertices). A coarse-to-fine approach to the shape recovery problem induces a fast sketch of the main vessels. Secondary vessels are extracted during a second phase (after a surface refinement).

Lachaud99a , Lachaud98b , Lachaud96a , Lachaud96c , Lachaud96d , Lachaud95 , Lachaud94