Abstract
Purpose :
To present a fully automatic method for the screening of three eye diseases in high-risk populations using information from retinal images, medical history and laboratory results.
Methods :
In the United States, diabetic retinopathy (DR), primary open-angle glaucoma (POAG) and age-related macular degeneration (AMD) affect 12.5 million people, and over 50 million are at high risk to develop those diseases. Early detection of AMD and POAG through screening can lead to interventions that delay late-stage advancement.
We developed a method to screen high-risk population for these major eye diseases which combines information from retinal images, medical history and laboratory panels. The risk factors based on each type of data are combined to give a final screening result.
Retinal Images: We combine global and local features tailored to each disease. For global features, we generalize our previously developed method for DR screening which is based on multi-scale filtering, color and texture information. These features are combined in a classifier of risk for one of the three eye diseases. In addition, we extract local features that present with each disease. For DR we detect bright lesions; for POAG we calculate cup-to-disc ratio and vascular nasal shift; and for AMD we detect drusen.
Patient Data: Patient information includes sex, age,diabetes duration, blood pressure and BMI.
Blood Panels: Panels included are: comprehensive metabolic panel, lipid and diabetes (HbA1c).
Results :
1000 images were used to build and train the algorithms and give baseline results for each disease. Lab data and retinal images were used to train the combination of image and patient information. Validation was performed on a total of 90 independent cases with the following distribution: N=16 with AMD only; N=31 with DR only; N=16 with POAG; and N=27 controls.
Results for each disease are as follows. DR: sensitivity/specificity improved from 90%/71% to 90%/91% by adding patient information and Lab data; AMD improved from 88%/54% to 94%/70%; POAG improved from 88%/67% to 100%/74%.
Conclusions :
The results show that combining retinal feature detection approaches with other patient information increases the predictive screening power for each disease individually and for all three diseases when combined. This gives evidence that targeted screening for at-risk populations can be clinically effective for AMD and POAG.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.