Abstract
Purpose :
To assess the accuracy of deep learning models to detect pre-perimetric and perimetric glaucoma from fundus photographs.
Methods :
We included 3,272 eyes of 1,636 subjects that had participated in the ocular hypertension treatment study (OHTS). A total of 66,721 fundus photographs and visual fields were previously examined by independent readers from the optic disc and visual field reading centers of the OHTS and glaucoma development was further confirmed by an endpoint committee. We included 45,379 reliable fundus photographs from OHTS participants; 41,298 fundus photographs of eyes without glaucomatous optic neuropathy (GON) and normal visual field, as defined by the OHTS study, and 4,081 fundus photographs of eyes with either apparent glaucomatous optic neuropathy (GON) without visual field abnormality (pre-perimetric glaucoma) or visual field abnormality (perimetric glaucoma). We then trained and validated a MobileNetV2 deep learning architecture using 85% of the fundus photographs and further re-tested the models using 15% held-out fundus photographs.
Results :
The area under the receiver operating characteristic curve (AUC) of the deep learning model in detecting glaucoma was 0.95 (95% confidence interval 0.93 - 0.96). The AUC was improved to 0.97 (0.96, 0.98) on re-testing fundus photographs of only pre-perimetric eyes (eyes with apparent GON). However, the AUC decreased to 0.88 (0.86, 0.89) when we tested the deep learning model using the fundus photographs of only perimetric eyes (eyes with abnormal visual field but without apparent GON).
Conclusions :
Deep learning models can detect pre-perimetric and perimetric glaucoma from fundus photographs with a high accuracy. Perimetric eyes without apparent GON had a higher tendency to be missed by the deep learning algorithms compared to eyes with apparent GON.
This is a 2020 Imaging in the Eye Conference abstract.