Image registration and subtraction for the visualization of change in diabetic retinopathy screening

Abstract

Diabetic retinopathy (DR) is a leading cause of blindness in people of working age in developed countries. DR is characterized by lesions of the retinal microvasculature. This paper describes a technique for visualizing the changes which occur in retinal pathology over a 1 year interval. Pairs of retinal images of the same eye were acuired 12 months apart within a DR screening programme. The images were normalized, registered and subtracted to generate a contrast-enhanced difference image displaying changes in retinal appearance. The normalization factor and registration error were quantified for 104 pairs of retinal images.

Introduction

Diabetic retinopathy (DR) is a complication of diabetes mellitus characterized by lesions of the retinal vasculature. Despite the availability of effective treatments, advanced DR is a leading cause of blindness in people of working age. In the United States, DR is the primary cause of new cases of legal blindness in adults aged 20–74 [1]. Approximately 0.4 million US adults have advanced, vision threatening DR [2]. The prevalence of DR is set to increase significantly due to what has been described as a diabetes epidemic. Population growth, aging, physical inactivity and increasing levels of obesity have contributed to the projected rise in the prevalence of diabetes from 2.8% in 2000 to 4.4% of the global population by 2030 [3]. It is, therefore, likely that the global incidence of sight threatening retinopathy will increase without effective methods for early detection and treatment.

Laser treatment has been of great benefit in reducing visual impairment [4], [5]. However, treatment is much better at preventing vision loss than restoring lost sight. Sight-threatening diabetic retinopathy often remains asymptomatic well beyond the optimal stage for treatment, especially in the case of proliferative disease. The efficacy of laser treatment is dependent upon pre-symptomatic detection of treatable pathology. With appropriate treatment at the optimal time, blindness can be prevented in at least one eye in over 90% of patients with proliferative retinopathy [6]. The prevalence of diabetes and the efficacy of existing therapies in the treatment of early stage diabetic retinopathy have prompted numerous public and professional organizations to recommend annual retinal evaluation for all patients with diabetes.

The UK National Screening Committee has recommended screening for diabetic retinopathy as a national priority [6]. The Northern Ireland DR screening programme (DRSSNI) currently screens approximately 12,000 patients per year with the majority of patients (75–80%) showing no significant DR. In fact, year on year many patients will continue to exhibit no significant DR. The retinal images from these patients must, of course, still be fully assessed after each screening episode to recognize that no significant change has occurred since the previous visit. This assessment typically consists of the visual scrutiny of the digital retinal image using limited image processing operations such as greyscale conversion and rescaling. The grading process is labor intensive with well trained graders able to assess images from approximately 20–30 patients per hour [7]. The rationale for this current proposal is that the majority of patients will have no detectable pathology at repeated time points (usually spaced annually). It should be possible for an individual to have a detailed grading of their retinal image during initial screening which establishes a baseline retinal appearance. A further image may then be acquired at the repeat screening episode 1 year later. The grader is interested in any images where change has occurred between each visit. These changes may be displayed using a difference image generated by subtracting the baseline image acquired during the first visit (V1) from the image acquired during the most recent visit (V2). If there is no change in the appearance of the retina between V1 and V2, then no lesions will be visible on the difference image. It is then assumed that the grading at V2 is the same as that at V1. These images will not require detailed grading.

Research efforts in ophthalmology have previously exploited image subtraction and registration for the analysis of time series images acquired during retinal angiography. Hipwell et al. [8] used rigid-body registration to align retinal images acquired during fluorescein angiography. The sequence of registered images was used to generate parametric images for quantifying retinal circulation. Recent work has attempted to develop registration schemes specifically for the alignment of retinal angiograms. Nunes et al. [9] describe a multi-scale registration technique based upon global translation and local elastic transformations. Image subtraction has also been previously investigated for images acquired during retinal fluorescein angiography. Subtraction was used to present differences between images recorded at different times [10]. Subtraction angiography has also been used to separate arterial from venous flow within the eye allowing identification of delayed venous filling [11]. These studies were limited by patient movement between each image acquisition and noise within the imaging chain. Image subtraction in both cases relied upon the manual registration of serial images using visually identified anatomical landmarks.

This paper describes the feasibility of implementing standard image processing algorithms (image registration, subtraction and contrast enhancement) for visualizing change in digital retinal images acquired 12 months apart within a clinical screening programme for DR. The aim is to identify change in retinal status over time as an indication of pathology requiring detailed grading. This work is a step towards the computer-based detection of patients with no change in their retinal appearance.