Computer-Aided Diagnosis of Diabetic Retinopathy

Kenneth W. Tobin and Edward Chaum

Low-cost, computer-aided systems for early detection of diabetic retinopathy could save tens of thousands of sight-years annually in the U.S. alone. We are developing a system that uses content-based image indexing and retrieval to address this issue.

The World Health Organization estimates that 135 million people have diabetes mellitus worldwide and that the number of people with diabetes will increase to 300 million by the year 2025.¹ Almost 20 million Americans currently have diabetes and an additional 16 million working-age adults have pre-diabetes or obesity or both and are at high risk for developing diabetes. Visual disability and blindness from diabetes have a profound socioeconomic impact. Diabetic retinopathy is the leading cause of new blindness in working-age adults in the industrialized world. It has been estimated that as much as $167 million dollars and up to 85,000 sight-years could be saved annually in the United States alone by improving the screening methods for diabetic retinopathy (DR).²

There is a significant need to develop inexpensive, broad-based screening programs for DR to reduce the economic and social consequences of vision loss from this disease. Treatment for DR is available; the challenge lies in finding a cost-effective approach with high sensitivity and specificity that can be applied to large populations in a timely manner to identify those who are at risk in the early stages of the disease. Through a National Eye Institute sponsored research project (R01-EY017065) we are developing a novel approach using content-based image retrieval (CBIR) to accurately describe and diagnose diabetic retinal disease from digital fundus images. The project leverages feature-based indexing and retrieval algorithms we have previously applied to the areas of manufacturing³ and geographic science.⁴ Our goal is to automate the detection and diagnosis of diabetic eye disease from digital images taken in a primary-care setting such as that shown in Fig. 1. The images are compared to a web-based repository and diagnosis is determined by retrieval and comparison with visually similar historical images in the system.

Fig. 1. A digital color fundus image reveals details about early eye disease, especially with regard to telltale lesions, basal laminar drusen in this case.

Our approach to diagnosing DR is built upon a proven idea that visually similar content in a population of images can be related through physical cause.³ For example images in a repository representing semiconductor wafer defects can be used to associate an errant manufacturing process with a query image through historical precedence encapsulated in the repository. Visual content is associated with image pixels, while historical precedence is associated with ancillary metadata. For semiconductors, metadata includes product type, manufacturing process flow data, process layer, etc. For human retinas, metadata includes age, gender, ethnicity, diagnosis, duration of disease, relevant laboratory data, and manifestations associated with pathology. To diagnose DR and other prevalent eye pathologies, we are attempting to prove that our visual content hypothesis generalizes to other image domains.

For electronic fundus imagery, we characterize visual attributes of the retina by first detecting important structures such as the vascular arcades, optic nerve, macula, and prevalent lesions as shown in Fig. 2.⁵ We define a coordinate system centered on the macula and then extract features that relate lesion types and distributions to the macula, the highly sensitive region of the retina responsible for detailed central vision. Unlike other methods that attempt to classify manifestations and pathology through rule sets or supervised training,⁶ our approach locates a statistically similar population in the repository based on an index derived from image features and infers manifestation and pathology from the image and metadata, resulting in an estimate of a probability of pathology based on extracted features. From a Bayesian point of view, we determine the conditional probability of a given pathology, ω_i, given an observation of visual features, ƒ, i.e., p(ω_i|ƒ.

Fig. 2. This red-free fundus image shows automatic detection of the vascular arcades (green), optic nerve, macula region (black), and lesion population (red). These lesions indicate soft drusen in a patient exhibiting age-related macular degeneration.

Although we are still in the early stages of this research, we have made significant progress in the characterization of retinal structures across a wide variety of digital fundus images representing many pathologies and associated disease manifestations.⁵ We are successfully localizing our retina coordinate system and currently researching relevant visual features of the retina topology and lesion populations. Our future work includes the expansion of our current data repository to incorporate data from additional DR studies, and the eventual determination of the efficacy of this CBIR approach through a clinical study with the University of Tennessee Health Science Center, Mid-South Telehealth Consortium. This network serves as a nationally recognized model for the implementation of telehealth through its conduct of health care delivery research, implementation of evidence-based medical practice in rural communities, and the development of a long-term sustainable model of regional health care delivery. The network integrates dynamic collaborations with community partners to successfully facilitate the delivery of needed services to underserved populations in the Delta region using new telemedical technologies. Our long-term goal is to improve eye health on a societal scale through lower cost and more efficient and timely diagnosis and referrals, access to expert diagnosis in underserved populations, and high-throughput methods to meet the growing need for screening in rapidly expanding at-risk populations worldwide.

We would like to acknowledge The Plough Foundation, Memphis, Tennessee, and Research to Prevent Blindness, New York, New York. This research is supported in part by National Eye Institute grant R01-EY017065. This paper was prepared by the Oak Ridge National Laboratory.

Authors

Dr. Kenneth Tobin is a Corporate Research Fellow and Group Leader of the Image Science and Machine Vision Group at the Oak Ridge National Laboratory, Oak Ridge, TN. He is a Fellow of SPIE, Associate Editor for the Journal of Electronic Imaging, and has organized and chaired many SPIE conferences and events.

Dr. Edward Chaum is the Plough Foundation Professor of Retinal Diseases, Professor of Ophthalmology, and Associate Professor of Pediatrics, Anatomy and Neurobiology and Biomedical Engineering at the University of Tennessee Health Science Center, Memphis, TN. He is the Director of the Retina Service at the UTHSC Hamilton Eye Institute.

References

1. A. F. Amos, D. J. McCarty, P. Zimmet, “The rising global burden of diabetes and its complications: estimates and projections to the year 2010,” Diabetic Med 1997;14: S57-85.
2. J. C. Javitt, et al., “Detecting and treating retinopathy in patients with type I diabetes mellitus: Savings associated with improved implementation of current guidelines,” Ophthalmology. 1991 Oct; 98(10):1565-73.
3. K. W. Tobin, et al., “Content-based Image Retrieval for Semiconductor Process Characterization.” Journal on Applied Signal Processing, 2002. 2002(7).
4. K. W. Tobin, “Automated Feature Generation in Large-Scale Geospatial Libraries for Content-Based Indexing.” Journal of Photogrammetric Engineering and Remote Sensing, 2006. 72(5).
5. K. W. Tobin, E. Chaum, V. P. Govindasamy, T. P. Karnowski, O. Sezer, “Characterization of the Optic Disk in Retinal Imagery using a Probabilistic Approach”, SPIE International Symposium on Medical Imaging, San Diego, California, Proceedings of SPIE, Vol. 6144, February 2006.
6. R. Pai, A. Hoover, M. Goldbaum, “Automated Diagnosis of Retinal Images Using Evidential Reasoning,” 15^th Int’l Conf. on Systems Engineering, Las Vegas, NV, Aug. 2002.

Submitted by: Edward Chaum, Kenneth W. Tobin, V. Priya Govindasamy, Thomas P. Karnowski, Image Science and Machine Vision Group

[an error occurred while processing this directive]