A fully automated comparative microaneurysm digital detection system.

Abstract

A fully automated digital image processing system, which provides an objective and repeatable way to quantify microaneurysms in digitised fluorescein angiograms, has been developed. The automated computer processing includes registration of same-eye retinal images for serial studies, cutting of regions-of-interest centred on the fovea, the detection of microaneurysms and the comparison of serial images for microaneurysm turnover. The microaneurysm detector was trained against a database of 68 images of patients with diabetes containing 394 true microaneurysms, as identified by an ophthalmologist. The microaneurysm detector achieved 82% sensitivity with 2.0 false-positives per image. An independent test set, comprising 20 images containing 297 true microaneurysms, was used to compare the microaneurysm detector with clinicians. The microaneurysm detector achieved a sensitivity of 82% for 5.7 false-positives per image, whereas the clinician receiver-operator-characteristic (ROC) curve gives 3.2 false-positives per image at a sensitivity of 82%. It is concluded that the computer system can reliably detect microaneurysms. The advantages of the computer system include objectivity, repeatability, speed and full automation.