An Implicit Inter-subject Shape Driven Image Deformation Model for Prostate Motion Estimation

This paper describes a novel approach for model based estimation of a dense deformation field utilizing an implicit representation of shape changes. Unlike existing methods based on the Point Distribution Model (PDM), the proposed method is not affected by an incorrect point correspondence which is a major limiting factor in practical applications of the PDM with clinical data. The proposed method uses regression between parametric representations of pelvic organs' shape and corresponding dense displacement field parameterized by the stationary vector field. The regression function is learned based on the training data sets including subjects with representative organ deformations, where the inter- and intra- subject correspondences are established via the log-Euclidean diffeomorphic formulation. The evaluation of the proposed method is conducted both on synthetic examples to provide systematic experimental evidence of correctness of the implicit shape representation for shape-driven prediction of the deformation field and, real MRI data to show accuracy in terms of deformation and prostate position prediction. The results show an increased robustness of the proposed framework in comparison to PDM approaches and suggest potential of its application for adaptive radiation therapy of prostate.