Statistical Shape Model: comparison between ICP and CPD algorithms on medical applications

Abstract

AbstractThe increasing availability of 3D anatomical models obtained from diagnostic images exploiting Reverse Engineering techniques allows the application of statistical analysis in the quantitative investigation of anatomical shapes variability. Statistical Shape Models are a well-established method for representing such variability, especially for complex forms like the anatomical ones. Not by chance, these models are widely used for medical applications, such as guiding segmentation of the diagnostic image and virtual reconstruction of incomplete anatomic region. The application of a statistical analysis on a set of shapes representing the same anatomical region essentially requires that shapes must be in correspondence, i.e. constituted by the same number of points in corresponding position. This work aims to compare two established algorithms, namely a modified version of the Iterative Closest Point and the non-rigid version of the Coherent Point Drift, to solve the correspondences’ problem in the construction of a Statistical Shape Model of the human cranium. The comparison is carried out on the models using the standard evaluation criteria: Generalization, Specificity and Compactness. The modified version of the Iterative Closest Point delivers a better Statistical Shape Model in terms of Generalization and Specificity, but not for Compactness, than the Coherent Point Drift-based model.