Mesh correspondence improvement using Regional Affine Registration: Application to Statistical Shape Model of the scapula

Statistical Shape Models (SSMs) are efficient tools in several fields and especially medical applications. In order to be meaningful, SSM construction requires a dense correspondence matching between a significant number of objects, which remains a very challenging task. Nonrigid registration is an essential element in the SSM construction pipeline that allows establishing correspondences between shapes and capturing their local variations. We introduce the Regional Affine Registration (RAR) that is based on object segmentation where each region is affinely registered to its counterpart. The point-based RAR process can be used to initiate a farther nonrigid registration process. Experiments are performed using fiducial metrics to compare three registration approaches: RAR, segmentation-based, and classical approach. As an application, we integrate RAR in the construction of an SSM of the human scapula. The three approaches are used to match correspondences between 85 scapulae and a validation of the SSM is carried out by evaluating its compactness, generalization and specificity. Multilevel B-Spline nonrigid registration parameter optimization is also investigated in a feedback sense using an initial version of the model. The initial SSM is used to generate in-correspondence shapes, giving ground truth to optimization experiments. The RAR method proved to initiate better the nonrigid registration which gave more accurate correspondence among synthetic and real database shapes. This was also reflected in the SSM validation tests. We show that performing the nonrigid registration in an iterative manner did not necessarily improve the final result.