Automated Image Detection of Retinal Pathology PDF

Preface

With the start of the 21st century, digital image processing and analysis is coming of age. Advances in hardware for capturing the minute detail in biological tissues such as the retina, and the unrelenting improvement in computational power in accordance with Moore's law, have provided the basis for mathematicians, computer scientists, and engineers to apply pattern recognition and image analysis in medical and biological applications. A better understanding of disease processes, which incorporate preclinical markers that identify people at risk combined with medical advances in diagnosis and treatment pave the way for improvement in health care generally and specifically for people with retinal disease such as that found in diabetes.

Globally the prevalence of diabetes mellitus is on the rise and with it the associated complications including retinopathy, heart disease, and peripheral vascular disease. Early detection of features often not directly discernible by clinical investigation has the potential to reduce the global burden of diabetes and cardiovascular disease. Although there are good public health reasons for screening certain populations or sub-populations, several factors need to be considered as outlined by the World Health Organization. These include the following: the disease is an important health problem, the natural history of the disease needs to be understood, there should be a detectable early stage, and treatment at the early stage should be more beneficial than treatment at later stages of disease. Diabetes and cardiovascular disease meet these criteria.

Diabetic retinopathy (DR) and heart disease are associated with changes in the characteristics of the blood vessels either in the retina, heart, or in the peripheral circulation. The retina is a tissue that is easily accessible and investigated. Signs of developing or current diabetic retinopathy and heart disease include changes in vessel diameter, occurrence of vessel tortuosity, new vessel growth, small enlargements of retinal capillaries referred to as microaneurysms, small and large hemorrhages and lipid exudates. Changes in either venule or arteriolar diameter have been associated with an increased risk of diabetes, hypertension, cardiovascular disease, and stroke. Even small increases in blood sugar levels, being below the accepted concentration for the diagnosis of diabetes, can affect the retina and lead to the presence of microaneurysms. To assess an appropriate number of people in the community, screening methods have to be economical, accurate, and easily performed. Therefore, automated assessment of preclinical or clinical signs associated with diabetic retinopathy and cardiovascular disease has been of great interest.

Engineering tools such as digital image processing combined with advanced machine learning allow identification and automated classification of features, lesions, and retinal changes in digital images of the retina. Objective diagnostic criteria for diabetic retinopathy progression have been available for 30 years, with the Early Treatment Diabetic Retinopathy Study providing a robust tool in 1991. Various permutations of this classification system have been proposed but essentially the grading system includes: minimal, mild, moderate, and severe nonproliferative retinopathy and proliferative retinopathy, both with or without macular involvement. Each of these stages is associated with the presence of particular pathological features. By combining different branches of engineering and medicine it has become possible to utilize today's technology and to contribute to the needs of physicians in providing optimal health care.

For any computer-based classification system to be successful the images need to be of adequate quality and resolution. The resolution of digital images can now exceed 10 megapixels and arguably surpasses the 35mm photographic standard. There has been work carried out in image preprocessing that addresses uneven retinal illumination, poor focus, or differences in background epithelium hue. The optic disc is the main feature in the retinal fundus. Numerous methods in image analysis such as using the measure of pixel intensity variance of a window size equal to the optic disc for locating the optic disc to applying the Hough transform and snake-based algorithms for identification of the boundary of the optic disk have been proposed. Automated microaneurysm segmentation was first applied to fluorescein-labeled images in the early 1990s and recently extended to nonmydriatic color retinal images. The latter is well suited for large population screening as it is noninvasive and economical. The literature on retinal vessel segmentation is numerous and varied. More work is needed, not only to characterize and compare algorithms, but also to improve algorithms to achieve better reliability in segmenting vessels. Segmentation of the vascular tree allows the determination of length, diameter, and coverage. It is also required to allow identification of lesions associated with diabetic retinopathy progression by removing the vessels as a confounder. Matched filters, local operators such as wavelet transforms, local gradients in the form of the first and second derivatives, neural networks that sweep over the whole image, vessel tracking algorithms, and many other approaches have been proposed. Local variation in vessel widths and vessel branching patterns can then be used to identify venous beading, vessel nicking, and the arteriolar-venous ratio.

Advances in automated image detection of retinal pathology require interaction and dialogue between practitioners from diverse fields. In this spirit, the contributors to this book include engineers, physicists, computer scientists, and physicians to provide a nexus between these fields. Each contributor is a recognized expert and has made significant contributions to the development of automated analysis of retinal images. There is much that has been discovered and proved to be effective in research institutes but has failed to make the transition to general clinical practice. Therefore many problems are yet to be solved.

This book is intended for researchers in the diverse fields of engineering, mathematics, and physics, as well as biologists and physicians. It discusses the epidemiology of disease, screening protocols, algorithm development, image processing, and feature analysis applied to the retina. Hopefully it inspires readers that automated analysis of retinal images is an exciting field, both for the physician and nonphysician alike, and one in which many developments have been made, but much more needs to be done if the products of our labor are to benefit the health of the community.

Chapter 1, by Jelinek and Cree, presents the general argument for automated image detection of retinal pathology. The field of diabetic retinopathy in terms of pathologies that affect the retina and those that are possibly identifiable by automated processing is emphasized. Prevalence and incidence of diabetic retinopathy are discussed with respect to the need for automated assessment. The chapter further outlines the use of automated retinal assessment as an indicator for disease elsewhere such as cardiovascular disease and peripheral vascular disease. A brief overview of work to date of automated image assessment of retinal disease is presented and opportunities for further development are highlighted.

Before embarking on applying image analysis techniques to retinal images, one needs to be aware of the clinical context. Chapter 2, by Worsley and Simmons, focuses on the pandemic of diabetes that is spurred by the reduction in physical activity and increases in energy-dense foods leading to an increase in obesity, insulin resistance, and Type 2 diabetes. The chapter discusses the importance of diabetic retinopathy and its contribution to blindness as well as its utility in providing a window on the natural history of diabetes in the population. The main focus is on how diabetic retinopathy is defined and classified, the disease process, the population distribution, and its prevention and screening for early changes.

Chapter 3, by Abr`amoff and Niemeijer, addresses optimization of retinal image digitization and detection of retinopathy. There are several approaches one can take, that is, decide whether there is disease / no disease or decide on progression. The chapter focuses on the frequency of occurrence of disease indicators in relation to a typical screening population and how research emphasis needs to focus on making automated detection methods more robust for clinical practice.

Chapter 4, by B¨acklund, concentrates on outlining the reasons why well-designed retinopathy risk-reduction programs need to be implemented on a large scale. Considerable efforts are required to address the importance and difficulty of achieving reliable early diagnosis of DR at a reasonable cost and to evaluate computer-aided diagnosis. Observations of a practitioner who has some experience with computeraided diagnosis on why systems for automated detection of retinal lesions have made little impact to date, are presented.

Chapter 5, by Osareh, is the first to provide an application of the nexus between engineering and medicine. The chapter concentrates on detection of the optic disc and retinal exudate identification. Color retinal images are classified to exudate and nonexudate classes following some preprocessing steps. The authors investigate K nearest neighbor, Gaussian quadratic, and Gaussian mixture model classifiers. The optic disc detection is based on color mathematical morphology and active contours. Implications for screening are discussed.

Cree, in Chapter 6, outlines the differences between microaneurysms and dot hemorrhages as part of the disease process and how their presentation can be used for automated detection and classification. Microaneurysms are usually the first sign of diabetic retinopathy and a positive correlation between the number of microaneurysms and disease progression has been reported. A historical background to automated detection of microaneurysms is followed by an overview of the standard approach to microaneurysm detection and extensions of this standard approach including application for population screening of color digital images.

Certain subtle vascular changes in retinal images are believed to predict the risk of disease elsewhere in the body. In Chapter 7, Cheung, Wong and Hodgson, develop the case that the arteriolar-venous ratio measured from retinal images is a good predictor of risk of diabetes, cardiovascular disease, and stroke as well as hypertension.

To automate measurement of features, such as the arteriolar-venous ratio, one must first reliably segment the blood vessels of the retina. In Chapter 8, Soares and Cesar, focus on the use of wavelet transforms to segment retinal blood vessels, discussing first the theoretical background of 1-D, 2-D continuous wavelet transforms, and the 2-D Gabor wavelet as a specific application that is suitable for blood vessel segmentation. Pixels are classified as either belonging to the vessel or not using a Bayesian Gaussian mixture model.

Chapters 9 (by Fritzsche et al.) and 10 (by Witt et al.) set the scene for analysis of the blood vessels including vessel diameter and branching angle, which can be indicators of pathology. Fritzsche et al. concentrate on how vessels can be detected using diverse models such as edge detectors, cross-sectional models, as well as algorithms that use intensity measures combined with thresholding, relaxation, morphological operators or affine convex sets to identify ridges and valleys. Following consideration of vessel segmentation the measure of vessel width is considered. Two potential issues arise, one being the orientation in which the width is measured and determination of the points at which to begin and end the measurement. Witt and co-workers concentrate on feature parameters associated with the geometry of vessel segments and bifurcations. One of the first parameters describing vascular geometry was the ratio of arterial to venous diameters. The geometry of bifurcating vessels may have a significant impact on hemodynamics of the vessel network and include the bifurcation angle, junction exponent, and tortuosity. These parameters are then discussed in the context of clinical findings such as hypertension, peripheral vascular disease, and ischemic heart disease.

The last chapter, by Yogesan et al., investigates the use of teleophthalmology in clinical practice. The chapter focuses on issues of image size and quality, health care costs, and early detection of diabetic retinopathy; and continues by considering combining teleophthalmology with automated assessment.