Purchase this article with an account.
K. W. Tobin, E. Chaum, M. D. Abràmoff, P. Govindasamy, T. P. Karnowski; Automated Diagnosis of Retinal Disease in a Large Diabetic Population. Invest. Ophthalmol. Vis. Sci. 2008;49(13):3225. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
There are presently 21 million diabetics in the United States and the number is expected to double by 2050. Diabetic retinopathy (DR) is the leading cause of blindness yet treatments can effectively preserve vision in patients with DR if the disease is diagnosed at an early stage. The medical community should be screening 400,000 patients for diabetic eye disease every week in the United States alone. By 2025 this number will exceed 1 million patients every day, worldwide. Through this research we are developing a diagnostic method that uses visual content of a retinal image archive to automatically estimate the presence and stratification of retinopathy. In this paper we present performance results for a population of 25,107 images.
Our method for diagnosing retinal pathology is based upon the indexing of visual content in an extensive archive of digital fundus images. For our purposes, visual content is derived from the statistical characterization of the population of dark and bright lesions that may exist in the macular region. Our approach uses a query-by-example method to retrieve images from the archive that exhibit similar characteristics to a query image. The retrieved population is used to estimate the presence of disease and disease stratification in a probabilistic framework. The analysis sequence includes image quality analysis, normalization, lesion segmentation and description, index generation, and probability estimation.
We present predictive diagnostic results for a population of 25,107 images collected from a DR screening program in the Netherlands. These images and data represent a type 2 diabetic population collected from 20 different screening centers in an anonymized manner in accordance with HIPPA guidelines. The images were collected from three camera types. Formats vary in size and field of view. Performance is presented for diagnostic accuracy, sensitivity, and specificity under confounding conditions of varying image quality, image format, camera type, and operator skill. Accuracy ranges from 80% - 98% depending on processing conditions.
This study represents the most comprehensive analysis of our method to date on an extensive set of diverse digital fundus images. Performance results support the feasibility of establishing acceptable methods for low-cost, high-throughout, automated screening and diagnosis for DR in a screening-like environment.
This PDF is available to Subscribers Only