A deep learning approach to automatic detection of early glaucoma from visual fields.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early detection of glaucoma is crucial for effective treatment and vision preservation.
• Standard diagnostic methods for glaucoma can be subjective and may miss early signs.

Purpose of the Study:

• To evaluate the efficacy of multi-scale spatial information from visual fields (VF) using a Convolutional Neural Network (CNN) classifier for early glaucoma detection.
• To compare the performance of the CNN model against traditional clinical metrics for discriminating between healthy and early glaucomatous eyes.

Main Methods:

• Two datasets of visual fields were analyzed, classifying them into control and early-glaucomatous groups.
• A custom-designed CNN was trained using a novel voronoi representation of VF data.
• Saliency maps were generated to visualize the regions influencing the CNN's classification decisions.

Main Results:

• The CNN classifier achieved high Average Precision (AP) scores, outperforming standard clinical decision measures like Mean Defect (MD) and square-root of Loss Variance (sLV) on one dataset.
• While performance varied slightly across datasets, the CNN consistently demonstrated strong classification accuracy.
• Saliency maps provided clinically relevant insights into the CNN's decision-making process for individual visual fields.

Conclusions:

• The proposed CNN model demonstrates superior performance in discriminating early glaucoma from normal visual fields compared to existing clinical measures.
• CNN-based classification, supported by saliency visualization, offers a potential tool to aid clinicians in the automatic and accurate diagnosis of early glaucoma.