To Örebro University

Purpose: Diabetic retinopathy (DR) is highly prevalent in the multi-ethnic and low socioeconomic population of South Western Sydney. This is a significant public health burden, demanding efficient and cost-effective diabetes retinal screening. The multi-centre South Western Eye and Diabetes Deep Learning Algorithm study, supported by an Ophthalmic Research Institute of Australia grant, aims to develop and validate a novel deep learning algorithm, capable of grading DR, for a new diabetes retinal screening service. This abstract evaluates the performance of the Diabetic retinopathy OCT Open souRce Artificial Intelligence (Doctor AI©SK) program.

Methods: Doctor AI analyses fundus photographs and optical coherence tomography (OCT) images simultaneously to grade diabetic retinopathy. It was trained with over 50,000 fundus photographs and 8000 OCT scans, using a combination of Australasian and publicly available datasets. As a screening tool, the algorithm's operating point was optimised for sensitivity and negative predictive value, and its performance reevaluated. Clinical validation is being undertaken in the recruited population at each study site.

Results: For the detection of diabetic macula oedema from OCT images, Doctor AI achieved a 96.8% sensitivity, 98.1% specificity and 97.5% accuracy. Area under the receiver operating characteristic curve was 0.996. For the diagnosis of DR from fundus photographs, Doctor AI achieved a 90.4% sensitivity, 96.8% specificity and 88.9% accuracy. Preliminary clinical validation revealed an accuracy of 89% and 85% for the OCT and fundus photograph modules of the algorithm, respectively.

Conclusion: Doctor AI is a unique deep learning algorithm capable of diagnosing DR and diabetic macula oedema with high sensitivity and accuracy.