Evaluation of deep convolutional neural networks for glaucoma detection

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early detection and diagnosis are critical for managing glaucoma and preventing vision loss.
• Deep convolutional neural networks (DCNNs) offer potential for automated analysis of medical images.

Purpose of the Study:

• To evaluate the diagnostic performance of DCNNs for glaucoma detection using color fundus images.
• To assess the impact of image quality and size on DCNN performance in glaucoma discrimination.
• To identify key image regions contributing to glaucoma diagnosis via heatmap analysis.

Main Methods:

• A retrospective study utilizing 3312 color fundus images (glaucoma-confirmed, suspected, and non-glaucomatous eyes).
• Three DCNN models were trained and evaluated for their discriminative ability.
• Analysis included varying image sizes, heatmap interpretation, and the effect of 465 poor-quality images.

Main Results:

• All three DCNNs achieved an area under the curve (AUC) of 0.9 or higher.
• DCNN performance was slightly better when distinguishing confirmed glaucoma versus suspected glaucoma.
• Optic disc area was identified as the most critical region for glaucoma discrimination by heatmaps.
• Poor image quality significantly reduced AUC by 0.1 to 0.2.

Conclusions:

• DCNNs demonstrate significant potential as a tool for detecting glaucoma and glaucoma-suspected eyes from fundus images.
• High-quality image preprocessing and selection are vital for optimizing DCNN discriminative performance.
• Further research can refine DCNNs for more robust glaucoma screening.