Abstract
Diabetic retinopathy (DR) is the damage to the retina and is a complication that can affect diabetes population. It is one of the most common causes of blindness worldwide. In any diabetic retinopathy screening programme or population based clinical study, a large number of digital fundus images are captured. These images are diagnosed by trained human experts, which can be a costly and time-consuming task due to the number of images they have to examine. Therefore, this is a field that would greatly benefit from the development of automated fundus analysis systems. It may potentially facilitate healthcare in remote regions and developing countries where reading expertise is scarce. The aim of this thesis to automatically analyse fundus images. The inherent variations in such images pose challenges for in-depth understanding on the presence of various DR signs. In this thesis, I first developed approaches for extracting retinal blood vessels and microaneurysms. Singular Spectrum Analysis (SSA) plays a key role to obtain the main structural features from the cross-sectional profiles of candidate objects. A vessel distribution map containing major prominent vessel fragments is constructed by combining SSA and local information. Based on the vessel distribution map, the full vessel network is then tracked and obtained. In microaneurysm detection, the cross-section profiles of candidate objects are filtered through SSA in order to extract a set of features for classification. Both detections have been tested on the publicly available datasets and further large sets of fundus images containing both pathological and healthy retinal photographs, demonstrating their effectiveness through various comparisons. The thesis further investigated human level understanding of digital fundus images for retinal disease detection. This part of work indicated that convolutional neural networks have great potential in developing automated fundus image analysis. Two independent large datasets are used to train two different DR grading systems based on American Academy of Ophthalmology (AAO) grading standards as well as UK National Screening Committee (NSC) grading standards to demonstrate the effectiveness and generic nature of such approach. This method was tested on very large scale sets of images. My method achieves a substantial high sensitivity and specificity of 96.20% and 94.60% respectively compared with previous work, if assuming human’s manual grading is correct. If no assumption on the correctness of human’s grading, high level of statistic agreement is also achieved between the automated system and human.