Artificial neural networks and spatial temporal contour linking for automated endocardial contour detection on echocardiograms: a novel approach to determine left ventricular contractile function

Abstract

This study investigated the use of artificial neural networks (ANN) for image segmentation and spatial temporal contour linking for the detection of endocardial contours on echocardiographic images. Using a backpropagation network, the system was trained with 279 sample regions obtained from eight training images to segment images into either tissue or blood pool region. The ANN system was then applied to parasternal short axis images of 38 patients. Spatial temporal contour linking was performed on the segmented images to extract endocardial boarders. Left ventricular areas (end-systolic and end-diastolic) determined with the automated system were calculated and compared to results obtained by manual contour tracing performed by two independent investigators. In addition, ejection fractions (EF) were derived using the area-length method and compared with radionuclide ventriculography. Image quality was classified as good in 12 (32%), moderate in 13 (34%) and poor in 13 (34%) patients. The ANN system provided estimates of end-diastolic and end-systolic areas in 36 (89%) of echocardiograms, which correlated well with those obtained by manual tracing (R = 0.99, SEE = 1.44). A good agreement was also found for the comparison of EF between the ANN system and Tc-radionuclide ventriculography (RNV, R = 0.93, SEE = 6.36). The ANN system also performed well in the subset of patients with poor image quality. Endocardial contour detection using artificial neural networks and spatial temporal contour linking allows accurate calculations of ventricular areas from transthoracic echocardiograms and performs well even in images with poor quality. This system could greatly enhance the feasibility, accuracy and reproducibility of calculating cardiac areas to derive left ventricular volumes and ejection fractions.