دانلود رایگان ترجمه مقاله الگوریتم های پردازش تصویر دیجیتال در رتینوپاتی دیابتی (نشریه الزویر 2009)

الگوریتم های پردازش عکس دیجیتال در رتینوپاتی دیابتی

Algorithms for digital image processing in diabetic retinopathy

Abstract

This work examined recent literature on digital image processing in the field of diabetic retinopathy. Algorithms were categorized into 5 steps (preprocessing; localization and segmentation of the optic disk; segmentation of the retinal vasculature; localization of the macula and fovea; localization and segmentation of retinopathy). The variety of outcome measures, use of a gold standard or ground truth, data sample sizes and the use of image databases is discussed. It is intended that our classification of algorithms into a small number of categories, definition of terms and discussion of evolving techniques will provide guidance to algorithm designers for diabetic retinopathy.

1 Introduction

Over the last decade, high resolution color digital photography has been recognized as an acceptable modality for documenting retinal appearance. Images are easily captured using a conventional digital camera back, attached to a retinal camera body designed to compensate for the optics of the eye. The digital format provides a permanent, high quality record of the appearance of the retina at any point in time. Electronic storage, retrieval and transmission are possible without loss of image quality.

One well recognized application for retinal digital imaging is within screening programs for diabetic retinopathy (DR). This disease is the commonest cause of blindness in people of working age, has an effective treatment available to prevent vision loss but is asymptomatic until late in the disease process. The UK National Screening Committee currently recommends annual screening for all diabetic patients aged 12 years and over, using digital retinal photography (www.nscretinopathy.org.uk). Images may be captured at a venue convenient to the patient’s home or work and data then transferred to a central location where they are read and interpreted by trained graders. Quality assurance must be an integral component of any screening programme, and as in breast screening programs, a high proportion of all images should be double read.

Population growth, an aging population, physical inactivity and increasing levels of obesity are contributing factors to the increase in the prevalence of diabetes. The global prevalence of diabetes is expected to rise from 2.8% in 2000 to 4.4% of the global population by 2030 [1]. In the UK the number of diabetic people is approximately 2.3 million (www.diabetes.org.uk). If all diabetic people are to undergo regular screening within a quality assured framework, the workload is going to be substantial.

Grading retinal images for the presence of diabetic retinopathy is largely a pattern recognition task. The typical features of diabetic retinopathy are microaneurysms, small intra retinal dot hemorrhages, larger blot hemorrhages, all of which are red lesions, and whitish lesions for example lipid exudates, and cotton wool spots which are nerve fiber layer microinfarcts. Graders are taught to recognize these lesions against the background appearance of the ‘normal retina’. With an increasing diabetic population and the need for quality assurance pathways, it is not surprising that considerable effort has been spent over the past 10–15 years on investigating whether these lesions could be detected by computer aided pattern recognition algorithms.

The process of detecting multiple patterns and their relationship within a retinal image is made up of a series of operations or steps, with low-level image processing operations providing a basis for higher level analysis. Digital retinal images are usually processed in an algorithmic sequence, with the output of one stage forming the input to the next. For example, a typical sequence may consist of one or more preprocessing procedures followed by image segmentation, feature extraction and classification stages. Preprocessing may be used to normalize image brightness, correct for image nonuniformity, reduce noise or reduce image artifacts. Segmentation decomposes an image into its constituent regions or objects, for example retinal blood vessels, optic nerve head or pathological lesions. Feature extraction typically computes quantitative infor mation from the segmented objects. The extracted features can be used to classify objects according to predetermined criteria such as size, morphology and color. The objectives of this paper are

(1) to review the relevant literature over a 10 year period in the field of digital image processing in DR;

(2) to provide researchers with a detailed resource of the main algorithms employed;

(3) to categorize the literature into a series of operations or steps;

(4) to identify potential areas for improving research design and reporting.

The paper is organized as follows: Section 2 describes the methodology used for the literature review. Section 3 gives the results of the review. Section 4 provides a detailed survey of the common computational steps for detecting retinal features. The image processing operations for detecting the optic nerve head, retinal vasculature, fovea, macula and retinopathy are described. We conclude in Section 2 by discussing recent trends and directions for future work.