Deep learning model to predict visual field in central 10° from optical coherence tomography measurement in glaucoma

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Visual field (VF) testing is crucial for diagnosing and monitoring glaucoma.
• Spectral domain optical coherence tomography (SD-OCT) provides detailed retinal layer thickness measurements.
• Predicting VF from SD-OCT could offer a non-invasive diagnostic alternative.

Purpose of the Study:

• To train and validate a deep learning (DL) model for predicting visual field (VF) sensitivity using SD-OCT data.
• To assess the prediction performance of the DL model compared to traditional machine learning methods.

Main Methods:

• A convolutional neural network (CNN) was trained on SD-OCT macular layer thickness data from 591 eyes (347 patients) with open-angle glaucoma (OAG) or normal subjects.
• Pattern-based regularization was used to prevent overfitting.
• Prediction accuracy was evaluated on an external dataset of 160 OAG eyes using absolute error (AE) and R-squared (R²) for whole VF and individual points, compared against Support Vector Machine (SVM) and Multiple Linear Regression (MLR).

Main Results:

• The CNN model achieved significantly lower mean AE for whole VF (2.84±2.98 dB) and point-wise AE (5.47±3.05 dB) compared to SVM and MLR (p<0.001 for both).
• The R² for the whole VF mean TH was 0.74 with CNN, and 0.44±0.24 for the 68 individual points.
• The DL model demonstrated superior prediction performance over SVM and MLR.

Conclusions:

• Deep learning models can accurately predict visual field sensitivity from SD-OCT scans.
• This approach shows potential for improved glaucoma assessment and management.