A generalised computer vision model for improved glaucoma screening using fundus images.

Area of Science:

• Ophthalmology
• Computer Vision
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early detection is crucial but challenging, especially in resource-limited settings.
• Computer vision models offer potential for enhanced glaucoma screening.

Purpose of the Study:

• To develop and validate a generalized deep-learning algorithm for glaucoma screening using fundus images.
• To assess the model's performance in distinguishing between healthy and glaucomatous fundus images.
• To evaluate the model's generalizability across diverse datasets.

Main Methods:

• Collected 18,468 fundus images from 20 public databases.
• Trained a deep learning model using Fastai and PyTorch on the aggregated dataset.
• Evaluated model performance using AUROC, sensitivity, specificity, accuracy, precision, and F1-score.

Main Results:

• The model achieved an AUROC of 0.9920 on the primary dataset.
• Sensitivity, specificity, accuracy, precision, and F1-scores exceeded 0.9530 for both classes.
• External validation on the Drishti-GS dataset yielded an AUROC of 0.8751 and accuracy of 0.8713.

Conclusions:

• The developed classification model demonstrates high efficacy in distinguishing glaucomatous from healthy discs.
• A slight decrease in accuracy on unseen data suggests the need for refinement and validation on larger, diverse datasets.
• The model shows potential for population-level glaucoma screening.