Neural Network and Fuzzy Clustering Approach for Automatic Diagnosis of Coronary Artery Disease in Nuclear Medicine

Abstract

We investigated feasibility of using fuzzy clustering and artificial neural network to predict coronary artery disease (CAD) in acute phase from planar and gated SPECT nuclear medicine images. We developed an automatic computerized scheme that helps physicians diagnose coronary artery disease based on 99Tc-Sestamibi myocardial perfusion images. Our study consisted of two separate studies with respect to patient population and imaging method. The first study included 58 subjects (30 male, 28 female) studies using planar rest and stress imaging and a second patient subset of 115 subjects (61 male, 54 female) using gated rest/stress SPECT imaging. After the myocardial perfusion scans, patients also had coronary angiography within three months of the imaging. Signal-to-noise ratio was improved by segmentation of myocardium from its background in both studies using fuzzy clustering with the Picard iteration algorithm. We extracted a set of adaptive features consistent with nature of nuclear medicine imaging and myocardium anatomy. Features were optimized and selected based on maximum separation in multidimensional feature space. A back-propagation artificial neural network (ANN) classifier was trained and tested for each study using the optimal features and the results of coronary angiographies as input and outputs, respectively. ANNs were trained, optimized, and tested using leave-one-out and Poh's Implementation of Weigned-Rumelhart-Huberman (PIWRH) methods, to diagnose the normal and abnormal patients based on their coronary angiograms. The performances of the optimal ANNs were analyzed by receiver operator characteristic (ROC) method. Results of ANN in the first study were compared to those of the physicians in nuclear medicine ward and two other physicians using ROC method. Results of ANN for the second study were compared to those of the nuclear medicine ward using ROC method. Both subsets demonstrate that the proposed method outperforms visual diagnosis and is therefore a useful adjunct for CAD diagnosis from planar and gated SPECT images.