Predicting 10-2 Visual Field From Optical Coherence Tomography in Glaucoma Using Deep Learning Corrected With...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate visual field (VF) testing is crucial for diagnosing and monitoring glaucoma.
• Predicting detailed HFA 10-2 VF tests from macular OCT measurements using deep learning (DL) shows promise.
• Integrating data from other standard tests like HFA 24-2/30-2 may enhance prediction accuracy.

Purpose of the Study:

• To evaluate if incorporating Humphrey Field Analyzer (HFA) 24-2/30-2 visual field (VF) data improves a DL model's prediction of HFA 10-2 VF tests.
• To assess the model's performance in predicting threshold sensitivity (TH) values from macular optical coherence tomography (OCT) measurements.

Main Methods:

• A convolutional neural network (CNN) DL model was trained using macular OCT data to predict HFA 10-2 TH values.
• The model incorporated pattern-based regularization (PBR) and was further refined by correction with HFA 24-2/30-2 data.
• Performance was compared between the CNN-PBR model alone and the model corrected with HFA 24-2/30-2, using absolute error (AE) and mean absolute error (MAE).

Main Results:

• The DL model incorporating HFA 24-2/30-2 correction demonstrated significantly lower AE (1.9dB vs. 2.6dB) and MAE (4.2dB vs. 5.3 dB) compared to the model without correction.
• Prediction errors were notably lower in the inferior temporal quadrant.
• The corrected model showed statistically significant improvements (P = 0.006 for AE, P < 0.001 for MAE).

Conclusions:

• Correction with HFA 24-2/30-2 data substantially enhances the predictive performance of DL models for HFA 10-2 VF from macular OCT.
• This approach optimizes the use of routinely collected HFA 24-2/30-2 and OCT data, potentially reducing the need for additional HFA 10-2 testing.