Machine Learning Approaches to Medical Decision Making

Recent advances in computing and developments in technology have facilitated the routine collection and storage of medical data that can be used to support medical decisions. In most cases however, there is a need for the collected data to be analysed in order for a medical decision to be drawn, whether this involves diagnosis, prediction, course of treatment, or signal and image analysis. Intelligent machine learning methods such as neural computing and support vector machines can be shown to be suitable approaches to such complex tasks. This thesis presents a study of the use of intelligent methods for medical decision making that aims to investigate and demonstrate their potential in such an application. The medical application presented in this work is the automated identification of tubercle bacilli (bacteria responsible for the tuberculosis disease) from photomicrographs of sputum smears. The development of such a system is of great importance since – according to the World Health Organization – tuberculosis kills more adults than any other infectious disease and its now on the increase. The diagnosis of tuberculosis is based on the screening of sputum smears under a microscope and therefore involves a medical image analysis task. Artificial neural networks and support vector machines have been employed for the detection and identification of tubercle bacilli and are supported by conventional image processing techniques. The basic concepts and algorithms underlying these methods are described and analysed in the first half of this documentation. The investigation of the main medical imaging and classification task that follows is carried out in two stages. First a preliminary study is described showing an overall generalization performance of up to 92.1% (achieving up to 93.5% sensitivity and 94.5% selectivity between different classification methods) on a small set of 1147 pattern examples from five different patients. This preliminary investigation has been used to support the fact that the development of a reliable and accurate classification system is feasible. A second study used larger data sets constructed from 65 Auraminestained smears. This investigation showed a slightly lower generalisation performance that reached an overall accuracy of 87.6% (achieving up to 93.9% sensitivity and 86.9% specificity between different classification methods). This second study involved several different data sets based on the medical preparation of the slides. Sputum smears used for identifying tubercle bacilli can be prepared using either Auramine/Rhodamine stain or a Ziehl-Neelsen stain yielding monochrome (fluorescent) or colour images. This presents an additional complexity to the image analysis and classification task, since both these modalities have to be taken into consideration. Although the set of smears and images used in the main investigation need to be much larger in order for a classification system to be considered statistically reliable, the results ii of this study continue to support that the development of such a diagnostic system is feasible. As concluded from the investigation a system trained on this data is still accurate enough to make a reliable decision supporting the diagnosis of the clinician. Different approaches, extensions and key points that can be used for improving the performance and use of a TB diagnostic system are discussed at the end of this documentation. Taking these suggestions under consideration, a system can be developed that can provide a greatly reliable and invaluable service to the medical community.