New synthesis of bayesian network classifiers and cardiac spect image interpretation

A new family of Bayesian network classifiers is introduced and demonstrated to outperform existing classifiers. Of particular interest is use of these classifiers for interpretation of cardiac SPECT images. High classification performance on databases from a variety of other domains is also demonstrated. Cardiac SPECT (Single Photon Emission Computed Tomography) is a diagnostic technique used by physicians for assessing the perfusion of the heart's left ventricle. A physician reaches the diagnosis by comparing SPECT images taken from a patient at rest and at maximum stress. Interpretation of images by strictly visual techniques is burdened with error and inconsistency. Thus, assistance in quantifying and automating the diagnosis is sought. An important issue in automating the diagnosis is classification of left ventricle perfusion into a number of predetermined categories. The goal of this dissertation is to investigate the use of Bayesian methods for construction of classifiers that would assist in interpretation of cardiac SPECT images. These images and their descriptions are characterized by a significant amount of uncertainty. Bayesian methods build models by approximating the probability distribution of the variables in the problem domain; they are naturally well suited to deal with uncertainty. This research consisted of three main parts. (1) Data warehousing—assembling cardiac SPECT images and patient records into an easily accessible database and creating software manipulation of SPECT images. (2) Three-dimensional image processing—implementation of custom algorithms for extraction of features from SPECT images. (3) Learning Bayesian network classifiers—research of novel machine learning algorithms that use Bayesian techniques for creation of robust classifiers. The main contribution of this work is creation of a new family of Bayesian network classifier—their high performance classifying left ventricular perfusion is demonstrated. Additionally, it is shown that they outperform existing Bayesian network classifiers and machine learning algorithm C4.5 using data from University of California at Irvine Repository of Machine Learning Databases. Among other contributions is a method for automated extraction of features from cardiac SPECT images based on the creation of models of normal left ventricles, software for visualization of cardiac SPECT images, automated feature extraction, and creation of Bayesian network classifiers.