Detection of diabetic retinopathy: a comparison between red-free digital images and colour transparencies

Abstract

The aim of this study was to compare how diabetic retinopathy was detected from red-free digital images and colour transparencies. Two ophthalmologists graded two-field, nonstereoscopic, 60 degrees red-free digital images and colour transparencies utilizing an ETDRS-based grading scale, from 107 mainly type 2 diabetic patients. The discordantly scored eyes were graded by the graders together to obtain a consensus level of retinopathy for each method. The eyes with discordant consensus grading results were further graded using all available photographic material to reach a final consensus level of diabetic retinopathy. Intermethod variations were presented as percentages and using kappa (k) and weighted kappa (wk) statistics. The errors of the two consensus gradings with respect to the final consensus grading were compared using McNemar's test. For the colour transparencies there was an agreement between the individual and the consensus grading results in 93% (k=0.90, wk=0.97) and 86% (k=0.79, wk 0.88) for grader 1 and grader 2. Corresponding figures for red-free digital images were 88% (k=0.83, wk=0.96) and 84% (k=0.78, wk 0.91). Agreement between methods was obtained in 76/107 eyes (71%; k=0.58 and wk=0.79). In the 31 discordantly graded eyes the level of retinopathy was underestimated in 20/31 (64%) vs 7/31 eyes (23%) and overestimated in 1/31 (3%) vs 3/31 eyes (10%) from colour transparencies and red-free digital images, respectively. The error tendencies were significantly lower when using red-free digital images (p<0.008). Red-free digital images are comparable with two-field colour transparencies in the identification of mild to moderate nonproliferative diabetic retinopathy.