Automatic differentiation of Glaucoma visual field from non-glaucoma visual filed using deep convolutional neural...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Glaucoma diagnosis relies heavily on visual field (VF) testing.
• Accurate differentiation of glaucoma from non-glaucoma VFs is crucial for timely treatment.
• Existing diagnostic methods have limitations in accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a deep neural network (DNN) for differentiating glaucoma from non-glaucoma visual fields.
• To compare the diagnostic performance of the DNN against human ophthalmologists and traditional scoring systems.

Main Methods:

• Collected 4012 visual field tests from 3 Chinese ophthalmic centers.
• Utilized Humphrey 30-2 and 24-2 test data, applying strict reliability criteria.
• Trained a Convolutional Neural Network (CNN) on 3712 VFs and validated on 300 VFs.

Main Results:

• The CNN achieved an accuracy of 0.876, with 0.826 specificity and 0.932 sensitivity on the validation set.
• CNN performance significantly outperformed ophthalmologists (average accuracy 0.585-0.626) and traditional rules (AGIS 0.459, GSS2 0.523).
• Traditional machine learning algorithms (SVM, RF, k-NN) showed lower accuracy (0.591-0.670) compared to the CNN.

Conclusions:

• The developed CNN-based algorithm demonstrates superior accuracy in differentiating glaucoma from non-glaucoma visual fields.
• This AI approach surpasses the diagnostic capabilities of human experts and conventional methods.
• The findings suggest a significant potential for AI in improving glaucoma diagnosis and patient care.