GONet: A Generalizable Deep Learning Model for Glaucoma Detection.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucomatous optic neuropathy (GON) affects millions globally, causing irreversible vision loss.
• Early detection is crucial but traditional diagnosis is time-consuming.
• Existing deep learning models for GON detection lack generalizability across diverse populations and settings.

Purpose of the Study:

• To develop a robust deep learning model for automated GON detection from color fundus photographs (CFPs).
• To enhance the generalizability of AI models for GON diagnosis across diverse datasets and patient demographics.

Main Methods:

• Developed GONet, a deep learning model using a DINOv2 pre-trained self-supervised vision transformer.
• Fine-tuned GONet using a multisource domain strategy on seven independent datasets comprising over 119,000 CFPs.
• Evaluated out-of-distribution generalizability and compared performance against state-of-the-art methods and cup-to-disc ratio.

Main Results:

• GONet demonstrated high out-of-distribution generalizability with an AUC of 0.88-0.99.
• Achieved performance comparable or superior to existing methods, improving cup-to-disc ratio assessment by up to 18.4%.
• A new open-access dataset of 747 CFPs with GON labels was contributed.

Conclusions:

• GONet offers a robust and generalizable solution for automated GON detection from CFPs.
• The model shows potential to improve early diagnosis and clinical management of glaucoma.
• The developed model and dataset are accessible to the research community to advance glaucoma research.