Dual-input convolutional neural network for glaucoma diagnosis using spectral-domain optical coherence tomography

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma diagnosis relies on detecting structural changes in the optic nerve head and retinal layers.
• Early detection of glaucoma is crucial for preventing irreversible vision loss.
• Spectral-domain optical coherence tomography (SD-OCT) provides high-resolution cross-sectional images of the retina.

Purpose of the Study:

• To evaluate the glaucoma-diagnostic capability of a deep-learning classifier using SD-OCT imaging.
• To assess the performance of a dual-input convolutional neural network (DICNN) in distinguishing glaucoma patients from normal subjects.
• To compare the diagnostic accuracy of the DICNN model with traditional convolutional neural network approaches.

Main Methods:

• A dataset of 777 Cirrus high-definition SD-OCT image sets (retinal nerve fibre layer and ganglion cell-inner plexiform layer) from 315 normal subjects and 462 glaucoma patients was curated.
• The dataset was divided into training and testing sets for model development and validation.
• A VGG16-based dual-input convolutional neural network (DICNN) was employed, utilizing both RNFL and GCIPL images as input for glaucoma diagnosis.

Main Results:

• The DICNN model achieved high accuracy in distinguishing glaucoma patients from normal subjects (accuracy=92.79%, AUC=0.957) on the test dataset.
• The model demonstrated significant diagnostic ability in identifying early-stage primary open-angle glaucoma (POAG) patients (accuracy=85.19%, AUC=0.869).
• The DICNN model outperformed traditional CNNs trained on individual retinal layers (RNFL or GCIPL) separately.

Conclusions:

• Deep learning algorithms utilizing SD-OCT imaging can effectively differentiate between normal individuals and glaucoma patients, including those with early-stage disease.
• The DICNN model, trained on both RNFL and GCIPL thickness maps, exhibits strong diagnostic performance for early glaucoma detection.
• This AI-driven approach holds promise for improving the accuracy and efficiency of glaucoma screening and diagnosis.