Development and Validation of a Deep Learning System for the Provision of a District-Wide Diabetes Retinal Screening Service.

Abstract

Background: Artificial intelligence (AI) enhanced retinal screening could reduce the impact of diabetic retinopathy (DR), the leading cause of preventable blindness in Australia. This study assessed the performance and validity of a dual-modality, deep learning system for detection of vision-threatening diabetic retinopathy (vtDR) in a multi-ethnic community.

Methods: Cross-sectional (algorithm-validation) study with the deep learning system assessing fundus photographs for gradability and severity of DR, and optical coherence tomography (OCT) scans for diabetic macular oedema (DMO). Internal validation of each algorithm was performed using a computer-randomised 80:20 split. External validation was by comparison to standard grading provided by two ophthalmologists in 748 prospectively recruited persons with diabetes (age ≥ 10) from hospital diabetes clinics and a general practice. Main outcome measures included sensitivity, specificity and the area under the receiver operating characteristic curve (AUC).

Results: Internal validation revealed robust test characteristics. When compared to ophthalmologists, the system achieved an AUC of 0.92 (95% CI 0.90-0.94) for fundus photograph gradeability, 0.91 (95% CI 0.85-0.94) for the diagnosis of severe non-proliferative DR/proliferative DR and 0.90 (95% CI 0.87-0.96) for DMO detection from OCT scans. It demonstrated real-world applicability with an AUC of 0.94 (95% CI 0.91-0.97), sensitivity of 92.7% and specificity of 95.5% for detection of vtDR. Ungradable images occurred in 55 participants (7.4%).

Conclusions: The dual-modality, deep learning system can diagnose vtDR from fundus photographs and OCT scans with high levels of accuracy and specificity. This could support a novel model of care for DR screening in our community.