Development and clinical deployment of a smartphone-based visual field deep learning system for glaucoma detection.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Glaucoma is a leading cause of irreversible blindness, projected to affect 100 million people by 2040.
• Current visual field (VF) testing methods for glaucoma diagnosis lack high efficiency.
• There is a critical need for advanced, accessible diagnostic tools for glaucoma detection.

Purpose of the Study:

• To develop and evaluate 'iGlaucoma,' a smartphone application-based deep learning system (DLS).
• To assess the diagnostic performance of iGlaucoma in detecting glaucomatous changes in visual fields.
• To compare the DLS performance against human ophthalmologists in glaucoma diagnosis.

Main Methods:

• A large dataset of 10,784 visual fields (VFs) from 5542 patients across seven Chinese centers was utilized.
• The study involved two phases: Phase I for training, validation, and testing the DLS, and Phase II for real-world cloud-based application testing.
• Diagnostic performance metrics, including area under the curve (AUC), sensitivity, and specificity, were evaluated against expert glaucomatologists.

Main Results:

• In Phase I, the DLS demonstrated superior performance compared to six ophthalmologists across three independent test sets (AUC 0.834-0.877).
• Phase II testing of the iGlaucoma cloud application showed high accuracy (0.99) in pattern recognition, with excellent AUC (0.966), sensitivity (0.954), and specificity (0.873).
• The system effectively detected glaucoma from Humphrey VFs, indicating strong clinical utility.

Conclusions:

• The 'iGlaucoma' smartphone application is a clinically effective tool for diagnosing glaucoma using visual field data.
• The deep learning system shows potential to enhance early glaucoma detection efficiency.
• Careful identification of the target population with expert input is recommended for optimal use of this diagnostic tool.