Purchase this article with an account.
Carla Agurto Rios, Sheila C Nemeth, Gilberto Zamora, Wendall Bauman, Peter Soliz, E Simon Barriga; Combining Medical Data and Fundus Images to Detect Eye Diseases in Patients with Diabetes. Invest. Ophthalmol. Vis. Sci. 2016;57(12):3404. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To integrate retinal images with other personal medical information, to develop a comprehensive eye evaluation algorithm that detects diabetic retinopathy (DR), age related macular degeneration (AMD), and risk for glaucoma in patients with diabetes, and also to show that this automatic detection can be improved when considering other medical information.
Studies have shown that diabetic patients are twice likely to have glaucoma and are at risk for AMD. Thus, screening programs that integrate glaucoma and AMD screening in their DR examinations are highly cost effective. Based in our previous work in automatic detection of DR, we expanded our system to detect these other two eye diseases. Our algorithm uses statistical features from retinal images and other medical information such as age, gender information, duration of diabetes, presence of hypertension and cataracts. Then, we combined these features using a partial least squares (PLS) classifier.To test our method, we used data from N=85subjects with diabetes. Of these, N=24 eyes had AMD, N=29 had DR, N=25 optic disc indicators of glaucoma (glaucoma suspect), and N=26 were controls, i.e., diabetics that did not present with any eye disease. In addition, we obtained medical history information such as age, diabetes duration, among others. The images were acquired using Canon retinal cameras. For each eye, optic disc-centered (field 1) and fovea-centered (field 2) images were used.
Table 1 shows the classification results in terms of AUC and sensitivity/specificity. Marked improvement was achieved for AMD classification, going from an AUC of 0.77 to 0.81 by adding the medical information. For glaucoma, we achieved a 3% improvement when adding medical features. For DR, image features were sufficient to achieve excellent classification accuracy (0.93).In general, the classification of combined eye disease also improved when the medical features are added. In this case, a 2% increase in AUC corresponds to 5% sensitivity from 84% to 89%, while the specificity remains in 64 %.
We present a new automatic system for the detection of the three main causes of visual impairment in the US: Glaucoma, AMD and DR. Our results show that the combination of retinal features with other patient health information has the potential to increase the performance of algorithms for automatic detection of eye diseases.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
Table 1 Preliminary results
This PDF is available to Subscribers Only