Justin Cooper , 2025-06-25 17:23:00
Key takeaways:
- The model had an area under the curve of 0.92 in the detection of severe glaucoma.
- Additional training data could boost the model’s performance in mild to moderate glaucoma.
MINNEAPOLIS — A deep learning model trained on fundus photographs showed promise in the detection of severe glaucoma, with lower accuracy in mild to moderate cases, according to a poster presented at Optometry’s Meeting.
“Glaucoma requires active screenings and calls for a reliable and affordable automated system to enable early detection, especially in high-risk communities,” Mikayla Kaliski, a third-year optometry student at the Southern California College of Optometry, said in a video presentation. “This leads to our project purpose: to leverage artificial intelligence to develop and evaluate the deep learning model for detecting the presence or absence of glaucoma from fundus photographs and categorize the stage of glaucoma as mild, moderate or severe.”
A deep learning model trained on fundus photographs showed promise in the detection of severe glaucoma, with lower accuracy in mild to moderate cases. Image: Adobe Stock
Kaliski and colleagues trained a model using electronic health record “data based on multimodal ophthalmologic investigation” matched with 6,665 fundus photographs from EyePACS, a diabetic retinopathy screening program. They assessed the model’s sensitivity, specificity, accuracy and area under the curve.
The model performed best in the detection of severe glaucoma, with 88% accuracy, 100% sensitivity and 83% specificity, with an area under the curve of 0.92.
As glaucoma severity decreased, so did sensitivity and area under the curve. Respectively, these were 48% and 0.62 for moderate glaucoma and 36% and 0.59 for mild glaucoma.
“Overall, the model shows promise for detecting severe glaucoma from retinal photographs,” Kaliski said. “To improve the model performance for detecting mild and moderate glaucoma stages, future work will involve incorporating additional source data to capture a diverse range of patient cases and increase the transferability of the algorithm, in addition to integrating [electronic medical record] data points such as patient demographic, family history and age into training the model to improve accuracy and reliability across all stages of glaucoma.”
Sources/Disclosures
Collapse
Kaliski M, et al. Development of a deep learning model for detecting mild, moderate, and severe glaucoma. Presented at: Optometry’s Meeting; June 25-28, 2025; Minneapolis.
Disclosures:
Healio could not confirm relevant financial disclosures at the time of publication.
