Optical Coherence Tomography Image Enhancement and Layer Detection Using Cycle-GAN.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Inconsistent optical coherence tomography (OCT) image clarity and retinal nerve fiber layer (RNFL) boundary delineation hinder glaucoma diagnosis.
• Deep learning, particularly generative adversarial networks (GANs), shows promise for medical image enhancement.

Purpose of the Study:

• To develop and compare deep learning methods for transforming low-clarity OCT images to high-clarity images from different devices.
• To concurrently estimate RNFL segmentation lines in the enhanced images.

Main Methods:

• Application of two deep learning models: pix2pix and cycle-GAN.
• Evaluation of image transformation using Fréchet Inception Distance (FID).
• Assessment of RNFL boundary delineation accuracy against manual annotations.

Main Results:

• Cycle-GAN achieved significantly lower FID scores, indicating superior image conversion compared to pix2pix (p < 0.001).
• Cycle-GAN demonstrated higher similarity to actual RNFL boundaries than pix2pix and manual annotations (p < 0.001).
• Enhanced images from cycle-GAN provided more precise RNFL boundary segmentation than original low-clarity scans.

Conclusions:

• Cycle-GAN is a more effective deep learning method for enhancing OCT image clarity and RNFL segmentation.
• The improved accuracy in RNFL boundary delineation holds potential for more reliable glaucoma diagnosis.