Creating realistic anterior segment optical coherence tomography images using generative adversarial networks

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Anterior segment optical coherence tomography (AS-OCT) is crucial for diagnosing ocular conditions.
• Generating high-quality synthetic AS-OCT images can aid in training AI models.

Purpose of the Study:

• To develop a generative adversarial network (GAN) for realistic high-resolution AS-OCT image synthesis.
• To evaluate the utility of generated images in machine learning tasks.

Main Methods:

• A Style and WAvelet based GAN (SWaN-GAN) was trained on 142,628 AS-OCT B-scans.
• Image realism was assessed using Fréchet Inception Distance (FID) and expert surgeon evaluation.
• A convolutional neural network (CNN) was trained with real and synthetic data for classification.
• Enhanced super-resolution GAN (ESRGAN) was used for high-resolution upsampling.

Main Results:

• Generated AS-OCT images showed high visual and quantitative similarity to real images (FID: 6.32).
• Surgeons could not reliably distinguish real from generated images (51.7% accuracy).
• CNN accuracy improved from 78% to 100% with synthetic data augmentation.
• ESRGAN upsampling yielded superior results (LPIS: 0.0905) compared to bicubic interpolation (0.4244).

Conclusions:

• GANs can generate high-definition, realistic synthetic AS-OCT images.
• Generated images are suitable for augmenting datasets in machine learning and image analysis.
• This approach holds potential for advancing AI-driven ophthalmic diagnostics.