Segmentation of Medical Images with a Shape and Motion Model: A Bayesian Perspective

This paper describes a Bayesian framework for the segmentation of a temporal sequence of medical images, where both shape and motion prior information are integrated into a stochastic model. With this approach, we aim to take into account all the information available to compute an optimum solution, thus increasing the robustness and accuracy of the shape and motion reconstruction. The segmentation algorithm we develop is based on sequential Monte Carlo sampling methods previously applied in tracking applications. Moreover, we show how stochastic shape models can be constructed using a global shape description based on orthonormal functions. This makes our approach independent of the dimension of the object (2D or 3D) and on the particular shape parameterization used. Results of the segmentation method applied to cardiac cine MR images are presented.