Primary Open-Angle Glaucoma Diagnosis from Optic Disc Photographs Using a Siamese Network.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Primary Open-Angle Glaucoma (POAG) is a major cause of irreversible blindness globally.
• Current deep learning methods for POAG diagnosis use single images, unlike clinical practice which compares baseline and follow-up images.
• Simulating the clinical comparison of images is crucial for accurate incident glaucoma diagnosis.

Purpose of the Study:

• To develop and evaluate POAGNet, a Siamese neural network designed to detect POAG from optic disc photographs.
• To simulate the clinical process of comparing baseline and follow-up images for glaucoma diagnosis.
• To assess the diagnostic performance and generalizability of the proposed POAGNet model.

Main Methods:

• Developed POAGNet, a Siamese neural network utilizing convolution to measure image similarity.
• Trained and evaluated POAGNet on two large datasets: the Ocular Hypertension Treatment Study (OHTS) and the Sequential fundus Images for Glaucoma (SIG) dataset.
• Incorporated deep supervision with two side outputs within the network architecture.

Main Results:

• POAGNet significantly outperformed state-of-the-art models on the OHTS test set (AUC 0.9587 vs. 0.8750) and SIG test set (AUC 0.7518 vs. 0.6434).
• The model demonstrated good generalizability, achieving comparable performance when trained on one dataset and tested on another (AUC 0.7490 on SIG from OHTS training).
• Training with combined OHTS and SIG datasets further improved performance (AUC 0.8165), indicating robustness.

Conclusions:

• POAGNet accurately diagnoses POAG by simulating the clinical grading process using optic disc photographs.
• The deep learning approach shows significant potential to assist and enhance clinical diagnosis of POAG.
• The POAGNet model is publicly available, facilitating further research and application.