Automated Construction of 3d Statistical Shape Models

Automated segmentation of medical images is a difficult task because of the complexity of anatomic structures, inter-patient variability, and imperfect image acquisition. Prior knowledge, in the form of pointbased statistical shape models (point distribution models) of a structure of interest can greatly assist segmentation to robustly find the structure in a patient’s image. Point distribution models are obtained through sets of corresponding landmarks lying on surfaces of training structures. The key to the automated construction of a three-dimensional (3D) statistical shape model is the identification of corresponding landmarks on training shapes, which is a challenging task. This paper presents a novel method for automated construction of 3D point distribution models. Corresponding surface points are obtained by two main steps: 1) volumes of interest (VOI), each containing one training structure, are manually defined, a reference structure is manually extracted from one training VOI and its surface is established and represented by a set of (reference) points, 2) reference landmarks are propagated to other training VOIs by transformations that are obtained by hierarchical elastic registration between the reference and each of the remaining training VOIs. We illustrate our approach using computed tomography data of the lumbar vertebra.