Diabetic retinopathy screening with computational support

Abstract

AbstractPurpose Diabetic retinopathy (DR) is among the most common causes of blindness in the developed world. Since many patients retain normal vision and may experience only minimal reduction despite the presence of a severe, sight‐threatening disease, screening and early detection of DR is highly recommended. Our purpose was to develop an automated DR screening system to gain computational support for high throughput screening activities.Methods Detection of microaneurysms (MA) ‐ the most important indicators and the earliest signs of DR ‐ and of other lesions (exudates, vascular structures, etc.) have been assembled in an ensemble‐learning based framework, which has been tested on 1200 images of the Messidor database.Results Images were classified as having DR or non by the MA detector only with 95% sensitivity, 51% specificity and 76% accuracy. By including other lesions, our system reached 89% sensitivity, 82% specificity and 85% accuracy.Conclusion Our system has sensitivity/specificity values for the MA step comparable to or better then other automated systems available, as tested in the Retinopathy Online Challenge, where it is currently ranked as first. MA detection remains a key component in automated DR screening, but detection of other DR lesions can lead to further improvement. This work was supported in part by NKTH, TECH08‐2, Hungary DRSCREEN project and the NIHR BMRC in Ophthalmology (TP).