Continuous Medial Representation
This project's goal is to extend and improve the medial modeling
methodology on which some of PICSL's morphometry and functional imaging
analysis projects are based. In the medial modeling technique,
deformable templates are fitted to anatomical objects under constraints
that preserve the geometrical structure of the template's
medial axis.
These constraints allow statistical morphometric analysis to leverage
the rich set of shape features by the medial axes. In particular,
medial axis explicitly describe the thickness of objects, and thickness
is a feature of great interest in studies involving neurodegeneration
and similar biological processes.
Deformable modeling with medial constraints presents a significant
methodological challenge. The boundary of a 3D object and its medial
axis are related by a set of non-linear equality constraints, which our
models must satisfy throughout deformation. Our present solution to
this problem is to define the medial model as a solution of Poisson
equation whose boundary conditions embody the non-linear constraints.
At present, this approach is limited to models whose medial axis is
formed by a single manifold. Such models have been used successfully to
answer hypotheses about the hippocampus in 3D
and the corpus callosum in 2D.
aims
Automatic Segmentation
Many anatomical structures
can not be delineated on the basis of intensity contrast with
surrounding tissues. Segmentation of such structures can benefit from
the Bayesian model-based approach, in which the shape and the
appearance of the target structure are modeled as probability
distributions. Medial models make it possible to use rich geometric
features like thickness in shape priors. Furthermore, medial models
offer an attractive way in which to sample appearance information:
image values inside of the object can be sampled along vectors
orthogonal to the model's boundary in a way that guarantees one-to-one
and onto parameterization of the object's interior.
Medial Branching Representation
Not all
structures can be modeled using a single medial manifold. We are
working on extending our method to modeling branching of 3D medial
manifolds. We indend to develop a model that circumvents having to
satisfy the non-linear constraints of medial geometry
in practice,
while still allowing the attractive properties of continuous medial
models to be leveraged. The initial application of this approach will
be to
modeling the myocardium.
bibliography