Abstract
Purpose :
To describe the development and validation of an artificial-intelligence based, deep learning algorithm (DLA) for the detection of referable diabetic retinopathy (DR).
Methods :
A cloud-based, convolutional neuron network (CNN) deep learning system for the automated detection of referable DR (≥pre-proliferative DR and/or diabetic macular edema) was developed and tested based on a set of 71,043 non-stereoscopic, retinal images. Retinal photographs were graded for DR severity by a panel of licensed ophthalmologists, with a gold standard grading for each image assigned when ≥3 consistent grading outcomes were achieved. For external validation, we also tested our DLA using 13,406 retinal images (any DR = 904; referable DR = 401) from three population-based cohorts of Malay, Caucasian and Indigenous Australians. Area under curve (AUC), sensitivity and specificity were the key outcome measures.
Results :
In the local validation dataset, the AUC, sensitivity and specificity of the DLA for referable DR was 0.989, 97.0% and 91.4%, respectively. Testing against the independent, multi-ethnic dataset achieved an AUC, sensitivity and specificity of 0.955, 92.5% and 98.5%, respectively. Eighty six percent of false positives cases were due to a misclassification of background DR. Undetected intra-retinal microvascular abnormalities accounted for 77% of all false negative cases.
Conclusions :
This artificial-intelligence based DLA can be used to detection referable DR from retinal images with high accuracy. This technology offers countless potential to increase the efficiency, accessibility and affordability of DR screening programmes.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.