A method for quantitative evaluation of statistical shape models using morphometry

We introduce a novel method for the evaluation of statistical shape models (SSM) that allows for quantifying the model quality wrt. global and local shape properties. The construction of SSM requires the identification of corresponding landmarks across a set of training shapes. Establishing such correspondence is a delicate matter and demands for automatic methods in a 3D setting. Conversely, the model quality needs to be evaluated to be able to compare different SSM in terms of specificity and generalization ability and to further improve the process of establishing correspondence. These well-known quantitative evaluation measures can be analyzed using various distance functions. The problem with popular landmark based metrics however is that the shape similarity of both the generated SSM and the actual object is disregarded. Evaluation of various models reveals that this can significantly corrupt the quality measures of the respective SSM, whereas the proposed method provides feasible results.