A Neural Network for the Detection of Glaucoma from Optic Disc Photographs

Abstract

Purpose : Optic disc photographs (ODPs) can be used to screen for glaucoma because they are easily attained and relatively effective to diagnose early disease. However, the requirement of trained interpretation reduces their utility in regions with limited healthcare. Automation of image assessment with artificial intelligence may help address these challenges. In this study, we propose a deep learning approach for the detection of early glaucomatous damage from a single ODP.

Methods : ODPs were graded independently by two glaucoma specialists as either “glaucomatous” or “healthy” based on the appearance of the optic nerve without reference to visual fields or the contralateral nerve. A total of 2265 ODPs were selected from the clinical database at UCLA (1660 total: 1610 glaucoma, 50 healthy) and the publicly available ACRIMA (297 healthy) and RIM-ONE (313 healthy) databases. Glaucomatous eyes were required to have mean deviation (MD) > -6 dB to focus training on early disease.

Images were split into training/validation/test sets of 1218/277/226 images. All images were automatically cropped at the disc margin, and all right eye images were flipped to left. The model used transfer learning with ImageNet weights to initialize the EfficientNetV2B0 model; it was then fine-tuned by unfreezing the last three layers of EfficientNet. Early stopping was enabled.

The model output a prediction score between 0 and 1 with the likelihood of the image being glaucomatous. An image with a prediction score of ≥ 0.5 was labeled glaucomatous.

Results : The model achieved an area under the receiver operating characteristic curve (AUC) of 0.95, overall accuracy of 0.92, sensitivity of 0.98, and specificity of 0.80 using grader labels as the ground truth.

Conclusions : Our neural network demonstrated clinically acceptable accuracy at detecting early-stage glaucoma based on a single ODP. It did so with a low false positive rate, which is important for limiting additional burden on existing health systems. This suggests that deep learning methodologies of ODP classification may be a useful tool in the diagnosis of early glaucoma to facilitate timely intervention and close monitoring, and that deep learning may be a viable option for population screening programs, particularly in under-resourced regions.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Confusion matrix of the model’s performance on the 226 test images (target refers to grader label)

 

Examples of images incorrectly classified by the model