Denoising diffusion-based anterior segment optical coherence tomography (AS-OCT) image generation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Annotated Anterior Segment Optical Coherence Tomography (AS-OCT) datasets are scarce in ophthalmology.
• Data imbalance issues hinder the development of predictive models.
• High-quality synthetic data is needed to train robust machine learning models.

Purpose of the Study:

• To generate synthetic AS-OCT images using Denoising Diffusion Generative Adversarial Networks (DD-GANs).
• To create diverse, realistic datasets for training predictive models without data imbalance.
• To address the scarcity of annotated AS-OCT data in ophthalmology.

Main Methods:

• Trained two DD-GAN models on healthy and unhealthy AS-OCT images from a tertiary referral hospital.
• Evaluated synthetic dataset quality using Fréchet Inception Distance (FID) and Inception Scores.
• Trained ResNet-50 models on real and synthetic data to compare performance.

Main Results:

• Generated two synthetic datasets (15.7k and 100k images).
• Achieved high-quality synthesis with low FID scores (0.17 healthy, 0.23 unhealthy).
• ResNet-50 models trained on synthetic data showed comparable performance to those trained on real data.

Conclusions:

• DD-GANs effectively generate realistic and balanced AS-OCT datasets.
• Synthetic data generation addresses ophthalmology data scarcity and imbalance, advancing medical image analysis.
• Synthetic medical image generation enhances data privacy by safeguarding patient confidentiality.