A Large-Scale Database and a CNN Model for Attention-Based Glaucoma Detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible vision loss.
• Automatic glaucoma detection using fundus images is an active research area.
• Existing methods struggle with high redundancy in fundus images, impacting accuracy.

Purpose of the Study:

• To propose an attention-based convolutional neural network (CNN) for enhanced glaucoma detection.
• To address the issue of high redundancy in fundus images for improved reliability and accuracy.
• To introduce a novel AG-CNN model incorporating attention mechanisms and pathological area localization.

Main Methods:

• Development of a large-scale attention-based glaucoma (LAG) database with 11,760 fundus images.
• Design of a novel AG-CNN architecture with attention prediction, pathological area localization, and classification subnets.
• Weakly supervised training using attention maps obtained from ophthalmologists via eye-tracking experiments.

Main Results:

• The proposed AG-CNN effectively highlights salient regions for glaucoma detection.
• Incorporation of localized pathological features enhances the model's performance.
• Experimental results demonstrate significant advancements over state-of-the-art methods on the LAG and a public database.

Conclusions:

• The AG-CNN model offers a significant improvement in automatic glaucoma detection.
• Attention mechanisms and pathological area localization are crucial for reliable glaucoma diagnosis.
• This approach advances the state-of-the-art in AI-driven ophthalmic diagnostics.