Developing Ultrasound-Based Computer-Aided Diagnostic Systems Through Statistical Pattern Recognition

Computer-aided diagnosis (CAD) is the use of a computer software to help physicians having a better interpretation of medical images. CAD systems can be viewed as pattern recognition algorithms that identify suspicious signs on a medical image and complement physicians' judgments, by reducing inter-/intra-observer variability and subjectivity. The proposed CAD systems in this thesis have been designed based on the statistical approach to pattern recognition as the most successfully used technique in practice. The main focus of this thesis has been on designing (new) feature extraction and classification algorithms for ultrasound-based CAD purposes. Ultrasound imaging has a broad range of usage in medical applications because it is a safe device which does not use harmful ionizing radiations, it provides clinicians with real-time images, it is portable and relatively cheap. The thesis was concerned with developing new ultrasound-based systems for the diagnosis of prostate cancer (PCa) and myocardial infarction (MI) where these issues have been addressed in two separate parts. In the first part, 1) a new CAD system was designed for prostate cancer biopsy by focusing on handling uncertainties in labels of the ground truth data, 2) the appropriateness of the independent component analysis (ICA) method for learning features from radiofrequency (RF) signals, backscattered from prostate tissues, was examined and, 3) a new ensemble scheme for learning ICA dictionaries from RF signals, backscattered from a tissue mimicking phantom, was proposed. In the second part, 1) principal component analysis (PCA) was used for the statistical modeling of the temporal deformation patterns of the left ventricle (LV) to detect abnormalities in its regional function, 2) a spatio-temporal representation of LV function based on PCA parameters was proposed to detect MI and, 3) a local-to-global statistical shape model based on PCA was presented to detect MI.