Development of a deep learning algorithm for automated diagnosis of retinopathy of prematurity plus disease

Abstract

Purpose: Blindness from retinopathy of prematurity (ROP) is largely preventable with timely detection and treatment. Treatment is predominantly indicated by the presence of plus disease, defined by posterior retinal vessel tortuosity and dilation. This study describes the development of a deep-learning algorithm, ROP.AI, able to automatically diagnose plus disease in fundal images. Methods: ROP.AI was trained using 6974 fundal images from Australasian image databases. Each image was assigned a diagnosis per real-world routine ROP screening and classified as nil plus or plus disease. The algorithm was trained using 80% of the images and validated against its remaining 20% within a holdout test set. Performance in diagnosing plus disease was evaluated within an independent set of 90 images. As a screening tool, the algorithm's operating point was optimised for sensitivity and negative predictive value, and performance re-evaluated. Results: Of 6974 fundal images, 5336 (76.5%) were graded as nil plus and 1638 (23.5%), plus disease. For plus disease diagnosis within the holdout test set, the algorithm achieved a 96.6% sensitivity, 98.0% specificity, and 97.3±0.7% accuracy. AUROC was 0.99. Within the independent set, the algorithm achieved a 93.9% sensitivity, 80.7% specificity and 95.8% negative predictive value. Following operating point optimisation, the algorithm diagnosed plus disease with a 97.0% sensitivity and 97.8% negative predictive value. Conclusion: ROP.AI is a deep-learning algorithm able to automatically diagnose plus disease with high sensitivity and negative predictive value. With ROP's growing global disease burden, its future development may allow for novel models of screening and care.L6320072812020-06-18 <br />