Convex spatio-temporal segmentation of the endocardium in ultrasound data using distribution and shape priors

We present a convex variational active contour model with shape priors, for spatio-temporal segmentation of the endocardium in 2D B-mode ultrasound sequences, which can be solved by Continuous Cuts. A four component (signal dropout, echocardiographic artifacts, blood and tissue) Rayleigh mixture model is proposed for modeling the inside and outside of the endocardium. The parameters of the mixture model are determined by Expectation Maximization, for the sequence. Annotated data is used to provide prior data, by which prior distributions for the inside and outside of the endocardium are constructed. Segmentation is then achieved by minimizing the Hellinger distance between prior and estimated distributions, under the constraints of a statistical shape prior built from principal eigenvectors of the annotated data. Since our model is convex, we can employ a fast optimization method: the Split-Bregman algorithm. Promising segmentation results and quantitative measures are provided.