Deep learning model to identify homonymous defects on automated perimetry.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Homonymous visual field (VF) defects often indicate serious intracranial pathology.
• These defects can be subtle and challenging to detect through traditional methods.

Purpose of the Study:

• To develop an automated deep learning artificial intelligence (AI) model for accurate identification of homonymous VF defects.
• To simplify the detection process of these critical neurological indicators.

Main Methods:

• Utilized Humphrey field analyzer (24-2 algorithm) data, processed via optical character recognition for mean deviation extraction.
• Developed a deep learning AI model (Deep Homonymous Classifier) using PyTorch and convolutional neural networks.
• Employed 7-fold cross-validation, data augmentation, and a complement cross-entropy loss function to train and enhance the model, addressing class imbalance.

Main Results:

• The AI model achieved an average accuracy of 87% in detecting homonymous VF defects on unseen data.
• Demonstrated a high recall rate of 92%, crucial for minimizing false negatives in disease detection.
• Achieved an F2 score of 0.89 and a Cohen's kappa value of 0.70.

Conclusions:

• The developed deep learning model is highly effective for identifying homonymous VF defects from automated perimetry.
• The AI model's performance indicates its potential as a valuable tool in clinical practice for early detection of neurological conditions.