Regional heart motion abnormality detection via information measures and unscented Kalman filtering

Abstract

This study investigates regional heart motion abnormality detection using various classifier features with Shannon's Differential Entropy (SDE). Rather than relying on elementary measurements or a fixed set of moments, the SDE measures global distribution information and, as such, has more discriminative power in classifying distributions. Based on functional images, which are subject to noise and segmentation inaccuracies, heart wall motion analysis is acknowledged as a difficult problem and, therefore, incorporation of prior knowledge is desirable to enhance the accuracy. Given noisy data and nonlinear dynamic model to describe the myocardial motion, unscented Kalman filter, a recursive nonlinear Bayesian filter, is devised in this study so as to estimate LV cavity points. Subsequently, a naive Bayes classifier algorithm is constructed from the SDEs of different features in order to automatically detect abnormal functional regions of the myocardium. Using 90x20 segmented LV cavities of short-axis magnetic resonance images obtained from 30 subjects, the experimental analysis carried over 480 myocardial segments demonstrates that the proposed method perform significantly better than other recent methods, and can lead to a promising diagnostic support tool to assist clinicians. © 2010 Springer-Verlag.