Spatio-temporal motion estimation for disease discrimination in cardiac echo videos

Abstract

In this paper we present a method of simultaneous registration of an entire sequence of frames of an echocardiographic sequence. In our approach, each echo frame is modeled using a probability density function, and registration problem between all pairs of echo frames is formulated as the problem of matching probability densities. An information-theoretic criterion called the Jensen-Renyi divergence is used to measure the distance between the probability density functions. The Renyipsilas Quadratic entropy results in a closed- form solution for the registration problem. Once the echo frames are registered, temporal trajectories of corresponding feature points in successive frames can be used to derive average velocity curves which have been shown to be useful for disease discrimination. To evaluate our technique for echo motion estimation for disease discrimination, we tested on a data set including cardiac echo from 21 patients of varying diseases. The data set includes a total of 72 complete cardiac cycles and contains 1612 frames. We compare our approach against two competing motion detection techniques, optical flow and Demons algorithm, on the same data set, and our motion detector performs best in terms of the separation between different diseases.