Assessment myocardial perfusion and contraction by Karhunen-Loeve transform on scintigraphic images.

Theory and previous studies showed that KLT (an application of principal component transform for imaging) can be use for analysis of cardiac function. This paper presents the results of our studies concerning the applications of KLT for images smoothing, quantification of myocardial contraction, and improvement of inter-observer reproducibility in cardiac imaging. The paper also describes the use of 4D cardiac phantom to quantify Karhunen-Loeve images. I. BACKGROUND CINTIGRAPHIC techniques are widely used to evaluate different parameters of cardiac function: myocardial perfusion, left ventricular ejection fraction, end-systolic and end-diastolic volumes, wall thickening etc. Quantitative analysis of the images sequences provides additional, clinically useful information. The Karhunen-Loeve transform (KLT) can be used for quantitative evaluation of cardiac function. The KLT is a statistical representation technique based on the changing of space that allows the description of a phenomenon, measured in a pattern (observation) space whose elements are correlated, by means of uncorrelated parameters in another space in which the principal axes are ordered in terms of importance (1)-(3). An important property of KLT is that, unlike the Fourier transform or the factorial analysis, the basic vectors are not known a priori but are "tailor made" for the given set of vectors. In the case of cardiac scintigraphic images, the original space is a sequence of images obtained by a discrete time sampling of the behavior of the object. Linear filtering of the sequence considered as a time function allows the extraction of the dominant information. The featured space is that of the eigenvectors of the covariance matrix of the initial data. These vectors are ranked according to a decreasing variance order. Significant parameters are therefore selected by keeping vectors associated with a maximum amount of variance. The images of KLT are obtained by projection of the initial sequence on the ranked eigenvectors. An