Purchase this article with an account.
Carla Agurto Rios, Eduardo S. Barriga, Gilberto Zamora, Victor Murray, Sergio Murillo, Honggang Yu, Wendall C. Bauman, Jr., Peter Soliz; Automatic Screening Of Eye Diseases Using 3-field Fundus Photographs. Invest. Ophthalmol. Vis. Sci. 2011;52(14):1342. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To present an automatic computer system that screens retinal photographs for the presence of lesions related to three major eye diseases: Diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma.
N=2247 retrospective digital fundus photographs were collected from the eyes of 822 patients at two eye clinics in San Antonio, TX: Retina Institute of South Texas (RIST) and University of Texas Health Science Center in San Antonio (UTHSC-SA). The photographs corresponded to three different fields of view: Disc centered, macula centered, and superior temporal. Not all patients had all fields in their set of images. Ground truth was provided by a certified ophthalmic medical technologist for the presence of pathologies including: microaneurysms, hemorrhages, exudates, neovascularization in the optic disc and elsewhere, drusen, abnormal pigmentation, geographic atrophy, and suspicion of glaucoma.A computer algorithm based on amplitude modulation-frequency modulation (AM-FM) and partial least squares (PLS) was used to produce binary results on the presence or absence of pathology.
The system achieved an area under the Receiver Operating Characteristic curve (AUC) of 0.89 for detection of non-Proliferative DR and of 0.92 for detection of sight-threatening DR (STDR). With a fixed specificity of 0.50, the system’s sensitivity ranged from 0.92 for all DR cases to 1.00 for clinically significant macular edema (CSME), using the presence of exudates within one disc-diameter of the fovea as a surrogate. For AMD pathologies, the system obtained a performance of AUC = 0.84 (sensitivity= 0.94, specificity= 0.50). For images that were flagged as glaucoma suspects by the human grader, the system achieved an AUC = 0.88 (sensitivity= 0.90, specificity= 0.88).
A computer-aided detection algorithm based on AM-FM and PLS was trained to detect different kinds of retinal pathologies. Of the DR pathologies studied, mild non-proliferative DR was the most challenging to detect, whereas the cases of STDR were detected with high AUC. Although the system was not originally intended for detecting abnormalities related to AMD or glaucoma, by adding some of those cases to the algorithm’s training database we were able to screen for these lesion with relatively high AUC. This work presents a viable and efficient means to characterize different retinal abnormalities and build binary classifiers for screening purposes.
This PDF is available to Subscribers Only