GONet: A Generalizable Deep Learning Model for Glaucoma Detection

AIMLab
Jun 18
2 min read

Introduction

Glaucoma is a leading cause of irreversible blindness, affecting over 60 million people worldwide. It damages the optic nerve and often progresses silently until significant vision loss occurs. Early detection is critical, but current screening methods are limited by their dependence on expert ophthalmologists and specialized equipment. Recently, deep learning (DL) models have been proposed to automate glaucoma detection from color fundus photographs (CFPs). However, these models often fail to generalize to new populations, devices, or clinical settings.

The Problem

Despite promising results, most DL models for glaucoma detection lack generalizability. They tend to perform well on curated datasets but struggle on unseen domains due to differences in patient demographics, camera types, or disease prevalence. Additionally, many existing studies rely on non-gold-standard annotations, where labels are based solely on image review rather than comprehensive clinical evaluations. These limitations hinder real-world adoption.

Our Contribution

In this study, we introduce GONet, a robust and generalizable DL model for glaucoma detection. The model is consists of a DINOv2 vision transformer, pre-trained via self-supervised learning and fine-tuned using a multi-source domain (MSD) strategy across seven independent datasets with gold-standard annotations. We also propose an end-to-end pipeline that includes:

1. Image quality filtering (FundusQ-Net)

2. Optic disc verification (LUNet)

3. Glaucoma classification (GONet)

Additionally, we introduce HYGD, a new open-access dataset with 747 CFPs labeled by glaucoma specialists.

Results

GONet achieved AUC scores between 0.88 and 0.99 across target domains - surpassing traditional features such as the cup-to-disc ratio (CDR) by up to 18.4%, and exceeding multiple published state-of-the-art results on target domains. It also demonstrated strong calibration with a Brier score of 0.08, indicating robust generalization.

*A figure showing the result of GONet vs. CDR and other baselines.*

Why It Matters

GONet addresses a major barrier in medical AI: poor generalization to real-world settings. By training across diverse populations with gold-standard annotations, it offers a more trustworthy and reliable tool for glaucoma screening. The model is freely accessible via Lirot.ai, enabling clinicians and researchers to evaluate and deploy the tool in practice.

*A figure showing a CFP, the optic cup and disc segmentation and an attention map showing where GONet focuses*

Conclusion

We introduce GONet, a deep learning model that achieves robust and generalizable glaucoma detection from CFPs. Validated across seven independent datasets, it outperforms existing baselines and contributes a new open-access dataset (HYGD) to the community. GONet represents a meaningful step toward accessible, AI-driven glaucoma screening in clinical environments.