GONet: A Generalizable Deep Learning Model for Glaucoma Detection

TL;DR

GONet tackles the poor cross-population generalizability of glaucoma detection from digital fundus images by leveraging self-supervised vision transformers pre-trained on large data and fine-tuned through multisource-domain strategies. Trained on seven diverse datasets with gold-standard annotations, GONet demonstrates strong out-of-distribution performance (AUC 0.85–0.99) and superiority over disc-based measures in most domains, while contributing HYRD as a new open dataset. The work systematically benchmarks SSL/vision-transformer configurations and shows that multisource-domain fine-tuning yields robust generalization beyond any single dataset. This approach enables reliable GON identification from a single fundus image, with potential for scalable, population-wide glaucoma screening and reduced reliance on specialized ophthalmologic assessments.

Abstract

Glaucomatous optic neuropathy (GON) is a prevalent ocular disease that can lead to irreversible vision loss if not detected early and treated. The traditional diagnostic approach for GON involves a set of ophthalmic examinations, which are time-consuming and require a visit to an ophthalmologist. Recent deep learning models for automating GON detection from digital fundus images (DFI) have shown promise but often suffer from limited generalizability across different ethnicities, disease groups and examination settings. To address these limitations, we introduce GONet, a robust deep learning model developed using seven independent datasets, including over 119,000 DFIs with gold-standard annotations and from patients of diverse geographic backgrounds. GONet consists of a DINOv2 pre-trained self-supervised vision transformers fine-tuned using a multisource domain strategy. GONet demonstrated high out-of-distribution generalizability, with an AUC of 0.85-0.99 in target domains. GONet performance was similar or superior to state-of-the-art works and was significantly superior to the cup-to-disc ratio, by up to 21.6%. GONet is available at [URL provided on publication]. We also contribute a new dataset consisting of 768 DFI with GON labels as open access.

Figures (7)

• Figure 1: Summary of the proposed research. We propose an end-to-end pipeline for the identification of GON from digital fundus images (DFI), based on a new deep learning model denoted GONet. Panel A: Low quality DFI filtration via FundusQ-Net Abramovich2023FundusQ-Net:Grading; Panel B: Missing OD filtration via LUNet Fhima2024LUNet:Images; Panel C: GON identification using GONet.
• Figure 2: Selection and labeling of the KULRD data. The flowchart summarizes the results of the data selection process.
• Figure 3: Result figures. Panel A: LUNet estimation of the vertical cup-to-disc ratio (CDR). Results are reported for the REFUGE dataset. Panel B: Performance comparison of alternative pre-trained vision transformers. Fine-tuning to the downstream task was performed on KULRD-Train and performance is reported for KULRD-Test. Panel C: Single-source domain training (SSD) versus multi-source domain training (MSD) strategies for a DINOv2 backbone. Panel D: Performance of GONet (DINOv2 backbone and MSD training) for GON identification versus baselines using the CDR or rim-to-disc ratio (RDR) and a benchmark open-source model called Brighteye Lin2024Brighteye:Transformer. For panels B-D, CI was calculated as detailed in section \ref{['subsec:performance_measures']}.
• Figure 4: Panel A: Examples of DFIs with low CDR ($\leq$0.5) that are GON+ (A.1) and DFIs with high CDR ($\geq$0.65) that are GON-, and which were identified as such with certainty of $\geq$0.75. Panel B: CDR distribution of GON+ and GON- DFIs per dataset. Panel C: Distribution of DINOv2 (C.1) and GONet (C.2) predictions for GON+ and GON- DFIs over all datasets, using kernel density estimation (KDE). Brier score Brier1950VerificationProbability is reported for each model.
• Figure I: GON+ (%): the prevalence of GON+ DFI in the original dataset (i.e., before applying exclusion criteria). #DFI Selected: the data subset used after applying exclusion criteria. FOV: the field of view of the fundus camera. Sex prevalence, age range are provided for the original dataset.