A Novel Statistical Piecewise Assembled Model (SPAM) for Reconstruction of Highly Deformable 3D Objects

Description

Deformable templates, and Point Distribution Model (PDM) in particular, provide a way of capturing and learning prior shape models and their variations from a training sample set in a stochastic way. This project proposes a novel PDM called 3D Statistical Piecewise Assembled Model (SPAM) that consists of a deformable global "frame" of the object shape and an assembly of piecewise parametric deformable surface segments. With this model, both the prototypical structure and the major manifolds of the 3D object and their modes of variations across samples of an object class can be learned and captured compactly with a small number of training samples. In addition, associated techniques of extracting context sensitive features for correspondence matching of 3D piecewise structures and automatic model landmarks selection will be investigated. Finally the feasibility of SPAM will be evaluated in biomedical applications such as the reconstruction of highly deformable soft tissue organs.