Purchase this article with an account.
E Simon Barriga, Carla Agurto, Honggang Yu, Vinayak Joshi, Cesar Carranza, Jeff Wigdahl, Sheila Nemeth, Gilberto Zamora, Wendall Bauman, Peter Soliz; FULLY AUTOMATIC RISK ASSESSMENT OF DIABETIC RETINOPATHY IN DIGITAL FUNDUS IMAGES. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5504.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To present a fully automatic system for diabetic retinopathy (DR) risk stratification using digital fundus photographs.
A software system was developed to automatically evaluate the risk of DR in digital fundus photographs. A block diagram of the system is given in Figure 1. The images are collected at a clinical site using a Canon CR-1 Mark II retinal camera. The imager collects at least one macula-centered image per eye. Optic-disc-centered photographs are also collected. Images are captured without mydriasis. Once the case is completed, the images are automatically uploaded to a PACS. The cases are automatically downloaded from the PACS and input to the DR Risk Analyzer™ (DR-RA). The processing steps are: First, the system evaluates the quality of the images and rejects those that are not of sufficient quality for image processing. Second, it automatically detects the optic disc and fovea in each image, its lateral orientation, and its alignment. Third all the valid images are then input to the risk analyzer, an algorithm based on amplitude modulation frequency modulation (AM-FM) and partial least squares (PLS). This algorithm outputs a continuous variable that represents the DR risk for that image. Finally, the DR-RA assigns a low- or high-risk determination to a case based on the analysis of the retinal photographs. A case is considered to be low risk if it does not present signs of DR or if it presents mild DR according to the International Clinical Diabetic Retinopathy Disease Severity Scale. A case is considered to be high risk if it presents moderate or severe NPDR, PDR, or macular edema.
The system was developed using 388 cases. When tested with N=99 independent cases, it achieved a sensitivity of 100% and a specificity of 79%. Further independent testing of N=478 cases (N=316 low risk and N=162 high risk) achieved a sensitivity of 95% and a specificity of 65%.
This work presents the development and testing of an automatic system for risk stratification of DR in fundus photographs. Mass DR screening programs would benefit from this system’s implementation by increasing the throughput of subjects that can be examined, and by making it accessible to remote and underserved populations.
This PDF is available to Subscribers Only