Multi-scale Multi-structure Siamese Network (MMSNet) for Primary Open-Angle Glaucoma Prediction.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Primary Open-Angle Glaucoma (POAG) is a leading cause of irreversible blindness globally.
• Current deep learning methods for POAG primarily focus on predicting current disease status using single images, not future onset.
• Glaucoma specialists assess disease progression by comparing serial optic nerve images with clinical data.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for predicting future Primary Open-Angle Glaucoma (POAG) events from fundus photographs.
• To simulate the clinical approach of comparing baseline and follow-up images for glaucoma assessment.
• To improve early detection of POAG for timely intervention and prevention of vision loss.

Main Methods:

• Proposed a Multi-scale Multi-structure Siamese Network (MMSNet) incorporating two-dimensional features and deep supervision.
• Trained and evaluated the MMSNet on a large dataset of 37,339 fundus photographs from 1,636 Ocular Hypertension Treatment Study (OHTS) participants.
• Conducted extensive experiments and an ablation study to validate the network's components and performance.

Main Results:

• MMSNet achieved superior performance on two "POAG prediction before onset" tasks compared to state-of-the-art methods.
• Achieved Area Under the Curve (AUC) values of 0.9312 and 0.9507, significantly outperforming existing approaches.
• Ablation studies confirmed the contribution of individual network components to the overall performance.

Conclusions:

• The developed MMSNet demonstrates significant potential for enhancing early prediction of future POAG events using deep learning.
• The findings highlight the efficacy of simulating clinical comparison of serial images for improved diagnostic accuracy.
• The proposed network offers a promising tool to assist clinicians in preventing irreversible blindness caused by POAG.