Single and Multiple Object Detection Problems in Medical Image Analyses
Balázs Harangi · University of Debrecen Electronic Archive (University of Debrecen) · 2015
In our days, diabetic retinopathy (DR) is the most common cause of blindness in the developed countries. In this PhD thesis, which is meant to be an answer to the question how we can improve the accuracy of screening this disease which affects a huge segment of the society, two basic components of an automated DR screening system are introduced. The first one is capable of locating the optic disc (OD) and the second one can detect exudates in the fundus images. Exudates are one of the first signs of DR, which appear on the fundus of the patient’s eyes. Thus, accurate and reliable detection of the exudates is an important part of a state-of-the-art computer aided diagnostic (CAD) system. We developed a novel algorithm which can detect this type of lesions with high accuracy. The basic idea is the following: a grayscale morphology-based approach extracts each of the possible exudate regions as a candidate extractor, then an active contour-based method determines the exact contours of these candidates. Finally, an optimally adjusted classifier selects the true exudates using region-wise features. As exudates appear in retinal images as bright patches, it is highly recommended that the OD should be localized and masked out before the detection of exudates would be started, as exudates are similar to the OD concerning color and shape characteristics. Thus, our aim is to determine the location of the OD with high accuracy. For this purpose, we propose an ensemble of individual OD detectors to improve their precision and show how the accuracy increases by using more information in order to localize the OD.