Session S94.3

Dense Motion Estimation of the Heart Using Gram-Charlier Based Mutual Information

M Rubeaux*, JC Nunes, M Garreau

Université de Rennes
Rennes, France

The goal of this work is to estimate the dense motion of the heart in Multi-Slice computed tomography (MSCT) imaging. MSCT offers new perspectives for cardiac kinetic evaluation with 4D dynamic sequences. It has become an important modality for physiological understanding and diagnosis of ischemic heart diseases, giving access to coronary vessels but also to the cardiac function. The development of robust 3D image registration methods is a mean to perform motion estimation and characterize cardiac function. However, the non-rigid motion of the heart and the large data volumes to process introduce additional difficulties in registration, and efficient new methods still have to be developed. Our method comes within the scope of mutual information (MI) based non-rigid registration techniques. MI has been extensively studied as a similarity measure for the registration of medical images, and has found to be robust. We propose a new way to compute the MI based on Gram-Charlier series which approximate probability density functions (PDFs). This method was primary developed in the context of Independent Component Analysis (ICA). Indeed, to overcome the problems induced by Parzen-like estimators, ICA researchers proposed to approximate the MI using higher order statistics. More precisely, the PDFs used to approximate entropies are computed using the cumulants up to fourth order. We then model non-rigid heart motion using cubic B-Splines, because of their computational efficiency, smoothness, and local control. In order to decrease computational time, we use a multi-resolution scheme coupled with a Broyden-Fletcher-Goldfarb-Shanno (BFGS) search procedure. We apply successfully our method to register cardiac MSCT volumes between successive times of the cardiac cycle. It gives access to a dense motion estimation of the heart. We have tested this method on both simulated and real images and showed the robustness and the gain of computational cost of our approach. We compared it to a range of MI algorithms and demonstrated that our method can overcome several problems inherent in the use of Parzen estimators. This method is generic enough to be applied to other cardiac modalities such as MRI in a future work.

(Abstract Control Number: 209)