Abstract
Purpose :
Diabetic retinopathy (DR), as sequela of diabetes mellitus, is a worldwide spreading disease which is forecasted to be strongly growing in the next years. Diverse telemedicine systems have been developed and evaluated for mass screening of DR, but almost all of these require an expert medical reader. We present an automatic computer system to detect DR using a software algorithm working on photographs avoiding expert readers.
Methods :
In a retrospective study, 406 fundus photographs of 219 patients attending the Department of Ophthalmology of the University Erlangen-Nuremberg were evaluated. Recording of images was performed with the fundus camera CenterVue DRS (Padua, Italy). For each image, a test version of the Integrated Tele-Ophthalmological System ITOS (Voigtmann, Nuremberg, Germany) automatically distinguished between three different stages: diagnosis compatible with DR, suspicion of DR, and DR not present. An ophthalmologist (first author) explored the retina clinically and performed a detailed anamnesis of all patients. In a second phase, pictures taken with the ITOS system were evaluated and classified by the same ophthalmologist. Finally, a correlation of both the ITOS system results and the retina expert was performed.
Results :
Using optimal settings, the operating characteristic (ROC) of the automated diabetic retinopathy red-lesion-detection displayed a sensitivity of 85.5% (95% confidence interval [CI] 74.2% - 93.1%) and a specificity of 90.9% (CI 87.3% - 93.8%).
Conclusions :
The automatic diagnosis of DR with a fully automatic computer system is suitable for mass screening in a population visiting an University Department of Ophthalmology. Sensitivity and specificity can be adjusted to the particular needs and values between 85% and 95% attained. The ITOS system allowed precise automatic early detection of DR, and bears the potential for a cost-efficient and effective screening method without the involvement of expert medical readers.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.