Knowledge based extraction of the left ventricular endocardial boundary from 2D echocardiograms

Abstract

Extraction of the endocardial boundary of the left ventricle is a key challenge in cardiac ultrasound imaging. The cardiac anatomy may be difficult to determine automatically without incorporating knowledge of both wall shape and intensity signature into the detection algorithm. The aim of this study is to establish a framework for knowledge based extraction of the left ventricular endocardial boundary. The method is based upon the Snake algorithm where internal and external energy terms are combined into a Snake energy. Instead of using the patient image directly for calculation of the external energy we propose to use the correlation between geometrically normalized images, from the patient and from the database. The ventricular shapes from the database cases are used to compute the internal energy term. One boundary is detected for each case, hence a selection criterion is required. The total Snake energy is evaluated for this purpose and compared to manual selection of the best case. As a preliminary verification of the framework, the ventricular end diastolic and end systolic areas and the ventricular ejection fraction were calculated from the detected boundaries for a set of patient cases, using both manual and automatic database case selection. Using manual case selection, the results are encouraging, but the total Snake energy did not provide a sufficiently robust selection criterion. The strength of the proposed method is its ability to utilize expert knowledge directly for extraction of the endocardial boundary from ultrasound data. Using manual selection of the best case, the calculated parameters from detected boundaries were in good agreement with manual delineation. Further work is required to find a robust selection criterion.