A Structure-Related Fine-Grained Deep Learning System With Diversity Data for Universal Glaucoma Visual Field Grading

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness globally.
• Current diagnosis and treatment are challenged by the lack of effective grading measures.
• Accurate visual field assessment is crucial for glaucoma management.

Purpose of the Study:

• To develop an artificial intelligence system for assessing glaucoma patients.
• To propose a fine-grained grading deep learning system (FGGDL) for visual field loss.
• To evaluate the system's ability to aid in glaucoma diagnosis and grading.

Main Methods:

• Collected 16,356 visual fields (VFs) from Octopus perimeters and Humphrey Field Analyzer (HFA).
• Developed FGGDL, a deep learning system, to evaluate VF loss and compared its performance to ophthalmologists.
• Discussed the relationship between structural and functional damage for comprehensive glaucoma evaluation.

Main Results:

• FGGDL achieved high accuracy (85-90%) and AUC (0.90-0.93) on HFA and Octopus data.
• The system outperformed medical students and performed comparably to ophthalmic clinicians (p=0.614).
• Cross-validation demonstrated improved diagnostic accuracy (p < 0.05).

Conclusions:

• A deep learning system (FGGDL) effectively grades glaucoma visual fields with high accuracy.
• The system's credible interface supports telemedicine and patient self-assessment.
• This AI tool offers a solution for adequate glaucoma patient assessment and management.