Abstract
Abstract: :
Purpose:To investigate the use of automatic image analysis for detecting diabetic retinopathy in retinal images captured with a digital non–mydriatic camera. Methods:82 patients (163 eyes) with either type 1 or type 2 diabetes mellitus were recruited at The Steno Diabetes Center based on medical records. All patients were photographed after the installation of mydriatic eye drops using a Topcon TRC–NW6S non–mydriatic digital fundus camera interfaced with a JVC 3CCD color camera with a pixel resolution of 1450 x 1026. Five overlapping non–stereoscopic 45° images (posterior pole, nasal, temporal, superior and inferior) were captured of each eye and stored in TIF format. The five images from each eye were arranged as a single mosaic image using an IMAGE net 2000 computer system, and two independent readers performed a masked grading using the ETDRS grading scale. In cases of disagreement a third retinal specialist served as an adjudicator. The images were also analyzed using commercial fundus image–analysis software (RetinaLyze System ®; RetinaLyze A/S, Hørsholm, Denmark). The system uses advanced modeling of the gray–level image function of digital images, primarily the green color channel, and provide automated red microaneurysm and hemorrhage lesion detection. The automatic image analysis was performed on the five individual images of each eye, and eyes with at least one detected lesion were classified as "Possible DR" by the system. Results:In total 42 eyes were graded as having no diabetic retinopathy (No DR or Questionable DR), and 121 eyes were graded as having diabetic retinopathy (Minimal DR or worse) by the human graders. A ROC curve demonostrating the range between sensitivity and specificity of the automated red–lesion–detection in detecting eyes with diabetic retinopathy was constructed: the AUC of the ROC curve was 90.8 % (95%–C.I.: 86.3 – 95.3). Setting the visibility threshold to 2.1 for the automated algorithm, the sensitivity of detecting eyes with diabetic retinopathy was 90.1 % and the specificity of detecting eyes without diabetic retinopathy was 73.8 %. The positive predictive value (PPV) was 90.8 % and the negative predictive value (NPV) 72.1 %. Conclusions:Our results indicate that automated detection of red lesions in digital fundus images may be used as a first–step screening tool for classifying diabetic patients with respect to the presence or absence of diabetic retinopathy demonstrating a potential for substantially reducing the burden of manual grading.
Keywords: imaging/image analysis: clinical • diabetic retinopathy • detection