Deep-learning Classifier With an Ultrawide-field Scanning Laser Ophthalmoscope Detects Glaucoma Visual Field Severity

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Accurate assessment of glaucoma visual field defect severity is crucial for timely intervention and management.
• Ultrawide-field imaging offers a broader view of the retina compared to traditional methods.

Purpose of the Study:

• To evaluate the diagnostic accuracy of a deep-learning (DL) classifier in detecting glaucoma visual field defect severity.
• To assess the performance of DL using an ultrawide-field scanning laser ophthalmoscope.

Main Methods:

• A total of 982 open-angle glaucoma (OAG) patients and 417 healthy individuals were included.
• Glaucoma patients were categorized into early, moderate, and severe stages based on Humphrey Field Analyzer 24-2 results.
• A deep-learning classifier was trained and tested on ultrawide-field scanning laser ophthalmoscope images.

Main Results:

• The DL classifier achieved a high area under the receiver operating characteristic curve (AUC) of 0.872 for differentiating normal eyes from all glaucoma patients.
• Performance varied by severity: AUCs ranged from 0.830 (early OAG) to 0.934 (severe OAG).
• High sensitivity and specificity were observed, particularly in distinguishing severe glaucoma.

Conclusions:

• Deep learning demonstrates high reliability in detecting glaucoma characteristics and visual field defect severity.
• Ultrawide-field scanning laser ophthalmoscopy combined with DL offers a promising tool for glaucoma assessment.
• This approach has the potential to aid in the early and accurate diagnosis of glaucoma.