Automatic detection of glaucomatous visual field progression with neural networks

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Accurate detection of visual field progression is crucial for managing glaucoma.
• Automated perimetry is a standard tool for visual field assessment.

Purpose of the Study:

• To evaluate the efficacy of computerized neural networks in detecting visual field progression in glaucoma patients.
• To compare the performance of neural networks against expert human observers.

Main Methods:

• A dataset of 233 Octopus G1 visual field series from 181 glaucoma patients was analyzed.
• Visual fields were independently assessed by three masked observers to determine progression.
• A backpropagation neural network was trained and tested on the visual field data to classify series as stable or progressed.

Main Results:

• The neural network achieved a sensitivity of 73% and a specificity of 88%.
• The network's classification concordance with expert observers was good (kappa range: 0.50–0.64).
• The model demonstrated reliable performance across multiple training and testing iterations.

Conclusions:

• Neural networks can be trained to accurately recognize visual field progression in glaucoma.
• The performance of the neural network aligns well with that of experienced clinical observers.
• This technology holds potential as a supportive tool for physicians in evaluating glaucomatous visual field changes.