Automated Interpretation of Myocardial Perfusion Images with Multilayer Perceptron Network as a Decision Support System

Abstract

Aim: Bull's eye pattern recognition with artificial neural networks (ANNs) has the potential to assist interpretation of myocardial perfusion images (MPIs). We aimed to develop a model for interpretation of MPI based on the clinical variables and imaging data. Materials and Methods: The study included 208 patients referred to the department of nuclear medicine for 2-day stress-rest ECG-gated MPI. Several ANN models were designed with the following input variables: average count of 20 segments of the bull's eye images of stress and rest MPIs, gender, the constellation of coronary artery disease risk factors and scintigraphic cardiac ejection fraction. The procedure was repeated excluding the data of the rest phase scan. Data of 150 subjects were used for training, 21 subjects for cross-validation and 37 subjects for final operation testing. Several ANN models were examined with different hidden layers and processing elements and functions. The target output variable was the conclusion of the nuclear physician (i.e., normal vs. abnormal scan). Results: A multilayer perceptron (MLP) with two hidden layers trained with both stress and rest data demonstrated the best performance to classify the normal and abnormal MPIs. It showed an overall accuracy of 91.9%, sensitivity of 91.3% and specificity of 92.9%. The accuracy of the similar MLP trained using stress-only myocardial perfusion images reduced to 67.6%. Conclusion: The automated interpretation of MPIs with a 2 hidden layer MLP trained with stress and rest images could be an accurate support system either for the interpretation or quality assurance.