Retinal Image Analysis for Eye Disease Detection and Classification

Omar, Mohamed (2018) Retinal Image Analysis for Eye Disease Detection and Classification. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) Omar.Mohamed_phd.pdf - Submitted Version Download (6MB) | Preview

Abstract

Diabetes is a chronic disease affecting over 2% of the population in the UK (Shaw et al., 2010). Diabetes of all types can lead to complications in many parts of the body, including the retina of the eye where it can lead to vision loss. In the retina, diabetes can lead to diseases called diabetic retinopathy (DR), diabetic macular edema (DME) and age-related macular degeneration (AMD) which are among the major causes of blindness in the working population of industrialised countries. Fortunately, diagnosing it in its early stages is useful for effective treatment and it is usually treatable. DR or DME are mainly characterised by red spots, namely microaneurysms and bright lesions, specifically exudates, whereas AMD is mostly identified by white or tiny yellow deposits known as drusen lesion. Since exudates might be the only early signs of DR, there is an increased demand for automatic DR diagnosis. Exudates and drusen may share a similar appearance; hence discriminating between them is of interest in enhancing diagnostic performance. However, due to the increased demand for automatic diagnosis from retinal lesion images, computerized systems have nowadays become essential. Although the process of diagnosis and decision-making is challenging for ophthalmologists when analysing retinal lesion images, it needs to be time-efficient. This thesis is concerned with an investigation of novel image processing techniques in the diagnosis of diabetes with different types of diseases such as DR, DME and AMD. The thesis focuses on feature extraction in image processing and proposes diagnostic techniques based on state-of-the-art technology. The first contribution of the thesis is the use of a new LBP-based feature extraction technique after dividing the image into number of patches instead of using the whole image in the analysis. The extracted features have been applied with both the radial basis function neural network (RBF-ANN) and k-nearest neighbour (k-NN) classifiers for training and testing purposes. The second contribution involves a novel technique used to propose a multi-label machine learning model for classification problem combining the diagnosis of diabetic macular edema (DME) with other factors such as patient age and LBP-based feature extraction being applied for classification. The third and main contribution of this thesis is to demonstrate that colour vector angles (CVAs) and local binary patterns (LBPs) are more suitable for the analysis and classification of retinal images and their use leads to high classification performance as compared to other textural and colour features. A codebook is generated using a bagged combination of local binary pattern (LBPs) and colour vector angles (CVAs) features to exploit colour and textural information.

The overall system has been assessed through intensive experiments using different classifiers with a dataset of 352 retinal images collected from various datasets, namely: DIARETDB0 (Kauppi et al., 2006), DIARETDB1 (Kälviäinen and Uusitalo, 2007), HEI-MED (Giancardo et al., 2012), STARE (Hoover and Goldbaum, 2003), and MESSIDOR (Decencière et al., 2014). Correct classification is reported with an average sensitivity of 99.51%, specificity of 99.60% and accuracy of 99.63% and an overall average area under the curve of 98.59%. This represents the best performance achieved so far when compared to existing state-of-the-art systems for the diagnosis of retinal lesions.

Actions (login required)

View Item

Downloads