Background: Diabetic retinopathy is the most common complication of diabetes mellitus and is one of the leading causes of vision impairment globally, which is also relevant for the Russian Federation.
Objective: To evaluate the diagnostic efficiency of a convolutional neural network trained for the detection of diabetic retinopathy and estimation of its severity in fundus images of the Russian population.
Methods: In this cross-sectional multicenter study, the training data set was obtained from an open source and relabeled by a group of independent retina specialists; the sample size was 60,000 eyes. The test sample was recruited prospectively, 1186 fundus photographs of 593 patients were collected. The reference standard was the result of independent grading of the diabetic retinopathy stage by ophthalmologists.
Results: Sensitivity and specificity were 95.0% (95% CI; 90.8-96.4) and 96.8% (95% CI; 95.5- 99.0), respectively; positive predictive value - 98.8% (95% CI; 97.6-99.2); negative predictive value - 87.1% (95% CI, 83.4-96.5); accuracy - 95.9% (95% CI; 93.3-97.1); Kappa score - 0.887 (95% CI; 0.839-0.946); F1score - 0.909 (95% CI; 0.870-0.957); area under the ROC-curve - 95.9% (95% CI; 93.3-97.1). There was no statistically significant difference in diagnostic accuracy between the group with isolated diabetic retinopathy and those with hypertensive retinopathy as a concomitant diagnosis.
Conclusion: The method for diagnosing DR presented in this article has shown its high accuracy, which is consistent with the existing world analogues, however, this method should prove its clinical efficiency in large multicenter multinational controlled randomized studies, in which the reference diagnostic method would be unified and less subjective than an ophthalmologist.
Keywords: diabetes mellitus; diabetic retinopathy; diagnostics; machine learning.; neural networks; screening.
Copyright© Bentham Science Publishers; For any queries, please email at epub@benthamscience.net.