Automated Computer System Grading of Diabetic Retinopathy from Fundus Photographs* Otomatik Bilgisayarlı Sistemle Gözdibi Fotoğraflarından Diabetik Retinopatinin Derecelendirilmesi

Purpose: To evaluate the efficacy of computer algorithms in the detection and grading of retinopathy derived from screening of diabetic photographs. Methods: Screening fundus photographs (5 per eye) taken with a mydriatic camera in a general ophthalmology clinic for 52 sequential diabetic patients were graded by a general ophthalmologist and a retina specialist and compared with computer grading of digitized images acquired from color slides. Results: Among the 104 eyes, 64 (61.5%) had no retinopathy lesions (Grade 10), 6 (5.8%) eyes had only 1 or 2 microaneurysms, 21 (20.2%) eyes had more than 2 microaneurysms but no other lesions (grade 21), and 18 (17.3%) had additional lesions (grade 21+). Among the ophthalmologists, there was exact agreement in the grading of 76 eyes; in 21 eyes grading by each was within 1 grade level, and in 7 the grades differed by 2 grade levels. Computer algorithm grading agreed with the ophthalmologists’ grade in 51 (50%) eyes; in 24 (23%) eyes the computer under-graded by one grade level, and by 2 grades in 2 eyes. In 22 (21%) eyes, computer grading was greater by one grade level than the ophthalmologists’ grading and in 5 (5%) eyes the algorithms overgraded by 2 grades. Conclusions: Computer algorithms were able to detect and grade retinopathy from fundus photographs taken of a set of sequentially screened diabetics with consistency similar to the agreement in grading between a general ophthalmologist and a retina specialist.