Posters

Aims: We describe a neural-network system for automated computerized screening of digital fundal images for retinopathy designed for two automated screening modalities: prefiltering, where software performs the initial check and refers onto a screener only 'suspicious'\udimages; and supported screening, where the software highlights and counts potential lesions, and assigns a grade, while allowing the user to reclassify lesions.\ud\udMethod: The system detects the optic nerve head, fovea, microaneurysms, haemmorhages and exudates, and distinguishes between exudates, drusen and light artefacts; it is trained on 8000 sample lesions/distractors from 80 randomly selected screening images. It diagnoses background retinopathy where there is at least one HMA present, and risk of maculopathy where there is an exudate within\udone disk diameter of the fovea.\ud\udResults: We evaluated 65 randomly selected test images. On a per lesion basis, the system has sensitivity/specificity of 85/90% for detection of HMAs; 86/86% for exudates. For background retinopathy (per-image) the sensitivity/specificity was 90/70%; for maculopathy\udrisk 90/91 %.\ud\udComments: False negatives for background retinopathy invariably have only one or two very small lesions. The system is thus suitable for prefiltering for background retinopathy, and supported screening for sight-threatening retinopathy, It can reduce the effort in manual screening by over 70%