Glaucoma detection in myopic eyes using deep learning autoencoder-based regions of interest

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma diagnosis in myopic eyes presents challenges.
• Current diagnostic methods rely on metrics like retinal nerve fiber layer (RNFL) thickness.
• Optical coherence tomography (OCT) provides detailed ocular imaging.

Purpose of the Study:

• To evaluate a deep learning autoencoder model for glaucoma detection in myopic eyes.
• To utilize regions of interest (ROI) from OCT texture enface images.
• To compare the model's accuracy against traditional methods.

Main Methods:

• A cross-sectional study included 453 eyes from 315 participants (healthy and glaucomatous).
• Swept-source OCT (SS-OCT) imaging was used to construct texture enface images.
• Four methods were compared: RNFL thickness, texture enface, single autoencoder, and dual autoencoder models.
• Diagnostic accuracy was assessed using Area Under the Receiver Operating Curves (AUROC) and Area Under the Precision Recall Curves (AUPRC).

Main Results:

• The dual autoencoder model achieved the highest AUROC (0.92) and AUPRC (0.86).
• This model significantly outperformed single autoencoder, RNFL thickness, and texture enface models (p < 0.05).
• No significant difference was found between RNFL thickness and texture enface measurements (p = 0.47).

Conclusions:

• The dual autoencoder model demonstrated superior diagnostic accuracy for glaucoma in myopic eyes.
• This deep learning approach, using ROI-based reconstruction error from OCT texture images, is a robust alternative to conventional metrics.
• The findings suggest potential for enhanced glaucoma classification using advanced AI models.