Abstract
The proposed multi-task 3D DL model demonstrated high generalizability in all the datasets and the presence of MF did not affect the accuracy of GON detection generally.
Introduction
Glaucomatous Optic Neuropathy (GON) is a leading cause of irreversible blindness. Clinicians typically diagnose it using Spectral-Domain Optical Coherence Tomography (SDOCT) to measure the thickness of the Retinal Nerve Fiber Layer (RNFL). However, myopia (nearsightedness) presents a significant hurdle, as it causes structural changes in the eye that mimic glaucoma on scans, leading to frequent false positives. This study introduces a 3D multi-task deep learning model designed to simultaneously detect GON and myopic features (MF), teaching the model to distinguish between the two conditions effectively.
Methods
The researchers developed a 3D deep learning model based on the ResNet-37 architecture. Unlike models that analyze 2D slices, this model processes the entire 3D SDOCT data cube (200 × 1000 × 200 pixels), capturing spatial relationships across the entire optic nerve head. The architecture uses a shared backbone for feature extraction that branches into two separate classification heads for GON and MF.
Datasets
This retrospective multi-centre study involved 8,151 SDOCT scans from 3,609 eyes. Training and internal validation data were collected from the CUHK Eye Center and Hong Kong Eye Hospital. The model was then tested on five independent external datasets from Hong Kong, the USA (Stanford University), and Singapore to ensure high generalizability across different populations.
Key Results
The 3D multi-task model demonstrated superior performance and robustness:
- Glaucoma Detection: In internal validation, the model achieved an AUROC of 0.949, significantly outperforming the standard clinical metric of average RNFL thickness (AUROC 0.913).
- Generalizability: In external testing across five diverse sites, the AUROC for glaucoma detection remained high, ranging from 0.890 to 0.950.
- Robustness to Myopia: Subgroup analyses showed that the model's performance for glaucoma detection did not significantly drop in the presence of myopic features, suggesting it successfully learned to ignore myopic structural noise.
Discussion and Limitations
The study highlights that learning to detect myopic features actually improves the model's ability to detect glaucoma by creating a more robust representation of ocular anatomy. However, limitations include the use of SDOCT en face images for MF grading rather than gold-standard color fundus photographs, and the current binary classification of myopia. Future work will focus on integrating quantitative data and testing the model in primary care screening settings.
Conclusion
The proposed 3D multi-task deep learning model offers a highly accurate and generalizable tool for glaucoma screening, particularly in populations where myopia is prevalent and often complicates traditional diagnostic methods.
Summary updated