Medical imaging algorithm research for diagnosis of ocular diseases

Color retinal fundus images provide visual documentation of the health of a person's retina. With the widespread adoption of higher quality medical imaging techniques and data, there are increasing demands for medical image-based computer-aided diagnosis (CAD) systems to manage large volumes of data, provide objective assessments for decision support and help in labour-intensive observer-driven tasks. This thesis focuses on the development of 2-dimensional color retinal image analysis algorithms for automated optic cup localization in glaucoma, the leading cause of irreversible blindness worldwide. Traditionally, the optic cup is automatically segmented using image processing-based methods, often with many hand-crafted heuristics. With the incorporation of learning based techniques, the accuracy of medical image-based CAD systems has improved significantly and are now widely accepted and adopted by medical practitioners. In this dissertation, three novel approaches for automatic localization of the optic cup in retinal fundus images are presented. In the first work, a boundary-based cup detection approach using vessel kinks is presented. The key contribution in this work is its close modeling relationship with the clinical grading protocol to identify the optic cup, providing explicit visual evidence. Experimental results demonstrated that the novel use of vessel kinks as cup boundary key points guidance provides improved accuracy performance over existing retinal image processing based strategies. Although the use of vessel kinks is highly desirable and provides additional visual evidence, accurate detection and interpretation of these small vessel bends can, at times, be challenging. Instead, in the second work, a novel region-based unsupervised learning approach for automatic optic cup localization is proposed. This approach requires no training procedure, and utilizes domain knowledge and region-based features in a similarity-based label propagation and refinement scheme to obtain an estimated cup region. The promising result suggests that learning-based techniques are capable of accurate automatic optic cup localization. Recently, supervised superpixel-based cup localization has demonstrated superior performance. In the third work, a study on the limitations of this state-of-the-art classification framework is presented and an alternative generalized multi-scale approach is proposed, with improved stability and performance. This approach offers a stable and robust solution to reduce classification performance variations due to repeated random sampling of training samples. Furthermore, it integrates and unifies multiple superpixel resolutions for better boundary adherence. Extensive experimental results demonstrates the improved robustness and accuracy in optic cup localization against existing methods. In summary, three approaches for optic cup localization are proposed. This thesis demonstrate that using vessel kinks as cup margin key points is highly desirable and provides additional visual evidence. The challenges in the detection of these key points are also discussed. Alternatively, a region-based unsupervised learning approach is presented. Experimentally, it was shown that in the absence of ground-truth labels, this approach is able to achieve higher or comparable accuracy to the boundary-based and existing retinal image processing-based approaches. Lastly, the limitations of the state-of-the-art supervised superpixel-based cup localization approach are studied and improved with a novel multi-scale multi-model framework, which offers stability and improved accuracy.DOCTOR OF PHILOSOPHY (SCE