Glaucoma classification in 3 x 3 mm en face macular scans using deep learning in a different plexus

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness globally.
• Early detection is crucial for managing glaucoma progression, but early stages are often asymptomatic.
• Broad screening methods are needed for early glaucoma detection.

Purpose of the Study:

• To evaluate computational features for automated glaucoma detection from retinal images.
• To compare the performance of deep learning-based features against handcrafted features.
• To assess the efficacy of these methods using a small field-of-view OCTA.

Main Methods:

• Utilized 3x3 mm en face optical coherence tomography angiography (OCTA) images.
• Extracted features using convolutional neural networks (CNNs) and traditional handcrafted methods.
• Classified images to differentiate between glaucoma patients and healthy controls.

Main Results:

• CNN-derived features performed comparably or superiorly to handcrafted features.
• CNNs outperformed handcrafted features on whole retina and superficial vascular plexus (SVP) projections.
• Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.967 on the SVP projection.

Conclusions:

• Automated feature extraction using CNNs shows high efficacy in glaucoma detection, even with limited imaging fields.
• CNNs demonstrate potential for improved early glaucoma screening.
• Attention map analysis suggests CNNs focus on diagnostically relevant retinal areas.