Deep learning models show efficacy in predicting glaucoma before its onset
Click Here to Manage Email Alerts
Deep learning models showed efficacy in predicting glaucoma development 4 to 7 years before disease onset, according to a study presented at the virtual Association for Research in Vision and Ophthalmology meeting.
“Glaucoma affects over 60 million individuals, over 3 million in the U.S. alone. Since, typically, there are no symptoms or signs at the early stages, 50% of individuals are unaware of the disease,” Siamak Yousefi, PhD, said.
Dilated fundus photography is currently the gold standard for glaucoma screening, but manual assessment of the optic disc requires significant clinical training and is labor-intensive and highly subjective. Recent advances in artificial intelligence, particularly deep learning (DL) models, may help glaucoma assessment, he said.
The aim of the study was to evaluate the utility of DL models for prediction of glaucoma from color fundus photography before the manifestation of clinical signs. From the large database of 66,721 fundus photographs of 1,636 subjects participating in the Ocular Hypertension Treatment Study, 41,298 fundus photographs were selected for training and testing DL models.
“We selected all fundus photographs from non-glaucoma eyes but only baseline and two follow-up visits of eyes that had eventually converted to glaucoma. We trained and validated a MobileNetV2 deep learning architecture using 85% of the fundus photographs and further retested the models using 15% held-out fundus photographs,” Yousefi said.
To mitigate the black box nature of the DL models and to verify regions that were most important for DL models to make a decision, activation maps were computed from fundus photographs of normal and glaucomatous eyes.
“Activation maps confirmed that optic cup and rim were the most important regions in the input fundus photographs,” Yousefi said.
The area under the receiver operating characteristic curve of the deep learning model in predicting glaucoma development 4 to 7 years before glaucoma onset was 0.77.
“Our proposed models may open new areas in developing deep learning models for predicting glaucoma more accurately, leading to identifying at-risk population. DL model outcomes may complement other routinely obtained medical examinations for glaucoma assessment,” Yousefi said.
Collapse
Thakur A, et al. Predicting glaucoma prior to its onset using deep learning. Presented at: Association for Research in Vision and Ophthalmology annual meeting; June 12, 2020 (virtual meeting).
Read more about
Click Here to Manage Email Alerts