Neural networks for visual field analysis: how do they compare with other algorithms?

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Early detection of visual field defects is crucial for timely intervention.
• Current algorithms for visual field defect detection have varying performance characteristics.

Purpose of the Study:

• To evaluate the performance of a feed-forward neural network in identifying visual field defects.
• To compare the neural network's performance against existing algorithms.

Main Methods:

• A single hidden layer feed-forward neural network was trained on visual field data from a glaucoma study.
• The trained network was tested on unseen data from the same study.
• Receiver operating characteristic (ROC) analysis was used for performance comparison.

Main Results:

• The neural network demonstrated higher sensitivity than global indices at >90% specificity.
• At lower specificities (80-85%), the neural network was less sensitive than cluster and cross-meridional algorithms.
• The cluster algorithm from the Low-Tension Glaucoma study showed significantly higher sensitivity at lower specificities.

Conclusions:

• Neural networks offer a potential tool for visual field defect detection, particularly at high specificity levels.
• Performance varies depending on the chosen specificity threshold.
• Further research may optimize neural network parameters for improved sensitivity across different specificity ranges.