Glaucoma Classification Using a NFNet-Based Deep Learning Model with a Customized Hybrid Attention Mechanism.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness globally.
• Early detection is critical as glaucoma often lacks symptoms and progresses silently.
• Peripheral vision loss is common, often unnoticed until central vision is impacted.

Purpose of the Study:

• To develop and evaluate a hybrid attention mechanism for enhanced glaucoma detection.
• To assess the performance of normalization-free ResNet architectures in conjunction with the attention module.
• To improve the accuracy and efficiency of automated glaucoma diagnosis from fundus images.

Main Methods:

• Proposed a hybrid attention mechanism to recalibrate feature maps for glaucoma classification.
• Utilized normalization-free ResNet (NF-ResNet) architectures (NF-ResNet-26, -50, -101).
• Evaluated the model on three public datasets (LAG, EyePACS, BrG) for differentiating normal from glaucomatous fundus images.

Main Results:

• The hybrid attention module with NF-ResNet architectures significantly outperformed state-of-the-art ResNet variants.
• NF-ResNet-50 with the attention module achieved high accuracy: 0.9394 (LAG), 0.9117 (EyePACS), 0.9020 (BrG).
• On a combined dataset, the model reached 0.9193 accuracy, 0.9182 sensitivity, and 0.9202 specificity.

Conclusions:

• The proposed attention module demonstrates exceptional performance in glaucoma detection.
• The hybrid attention module combined with normalization-free architectures is a highly competitive classification model.
• This approach offers a promising tool for early and accurate glaucoma diagnosis, aiding in vision preservation.