A Vision Transformer Model for the Detection of Glaucoma from Optic Disc Photographs

Purpose: This study uses a deep learning algorithm to analyze optic disc photographs (ODPs) and classify eyes as “glaucomatous” or “healthy” based on optic nerve appearance. Methods: ODPs from three databases were independently graded by two glaucoma specialists. Images were preprocessed using the open-source language R and RimNet, a deep learning model for optic disc segmentation, to prepare inputs for training. The model was developed with Python based on Google’s Vision Transformer (ViT). After assessing the ODPs, the model provided an output between 0 and 1 to predict glaucoma likelihood (≥ 0.5 signified glaucoma). Model performance was evaluated using the area under the receiver operating curve (AUC), where 1 indicates perfect classification. Results: A total of 1,610 glaucomatous eyes with a mean MD of -2.04 were analyzed using the model. The model achieved AUCs of 1.00 , 0.98, and 1.00 in the training, validation, and test phases respectively. Overall accuracy in test images was 0.987, sensitivity was 0.994, and specificity was 0.969 with grader labels as the ground truth. A later assessment using 956 severe glaucomatous eyes with a mean MD of -11.71 reached 99.9% accuracy. Conclusions: The model demonstrated high accuracy in detecting glaucomatous optic nerve damage from ODPs. The model’s strong potential for early disease detection suggests that deep learning can be a valuable, cost-effective tool in glaucoma screening, especially in resource-limited regions. Translational Relevance: Our study demonstrates the potential of deep learning in providing accessible, early-stage glaucoma detection, supporting global efforts to prevent vision loss.