Improvement of Statistical Shape Models for Non-rigid Tissues Using Coherent Point Drift Algorithm

Statistical Shape Models are used to interpret shapes. They include mean and variance of corresponding points of training shapes. One of the most important challenges in building statistical shape models is to establish correct correspondences among landmarks in a training set. In this paper, the non-rigid CPD (Coherent Point Drift) method is used to find correct correspondences among points. This method uses both Deterministic Annealing and a non-rigid scheme to register two shapes simultaneously. Then, the statistical shape model is built using a rigid transformation. The proposed method is evaluated using Compactness, Generalization ability and Specificity measures. The built model is compared to models created using the ICP (Iterative Closest Point), TPS-RPM (Thin Plate Spline – Robust Point Matching) and MDL (Minimum Descreption Length) methods by these metrics. The results show that the proposed method performs like the MDL regarding Specificity measure (0.21±0.06). The Compactness and Generalization ability measures of the proposed method are very similar to those for the MDL method. The run-time of our proposed method is about 68 seconds which is faster than non-rigid TPS-RPM and MDL approaches (390 and 3600 seconds respectively). Our results are superior to the ICP and TPS-RPM algorithms.