Researchers recently developed a deep learning system that analyzes retinal fundus images with a high sensitivity, specificity and generalizability for detecting glaucomatous optic neuropathy (GON).
This cross-sectional study evaluated 241,032 fundus images from 68,013 patients. The team tested the AI in several validation datasets, which allowed them to assess the tool in a clinical setting without exclusions—it also helped them test against fundus photographs of variable image quality.
The investigators found that 12.4% of the images showed definite GON, 4.6% showed probable GON and 83% showed unlikely GON. They note that the deep learning system model had a sensitivity of 96.2% and a specificity of 97.7% in the primary local validation datasets. They add that the most common reason for false-negative and false-positive grading by the artificial intelligence program (46.3% and 32.3%, respectively) and manual grading (44.2% and 34.0%, respectively) was pathologic or high myopia.
These findings suggest that use of AI for glaucoma “could enhance and expedite screening in a cost-effective and time-efficient manner,” the study authors conclude.
|Liu H, Li L, Wormstone M, et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA Ophthalmol. September 12, 2019. [Epub ahead of print].|