RF-GANs: A Method to Synthesize Retinal Fundus Images Based on Generative Adversarial Network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness, necessitating accurate and timely diagnosis.
• Current deep learning methods for DR detection are hindered by imbalanced datasets, specifically a lack of diverse retinal fundus images.
• Existing diagnostic tools require significant improvements to overcome data limitations.

Purpose of the Study:

• To propose a novel generative adversarial network-based method (RF-GANs) for synthesizing realistic retinal fundus images.
• To address the challenge of data imbalance in training deep learning models for diabetic retinopathy detection.
• To improve the performance of diabetic retinopathy grading models through data augmentation with synthesized images.

Main Methods:

• Developed a two-stage generative adversarial network: RF-GAN1 for domain translation and RF-GAN2 for image synthesis.
• RF-GAN1 translates retinal images between datasets to reduce domain gap, enhancing semantic segmentation model training.
• RF-GAN2 synthesizes new retinal fundus images using extracted masks and diabetic retinopathy grading labels.

Main Results:

• RF-GAN1 effectively narrows the domain gap, improving segmentation model performance.
• RF-GAN2 successfully synthesizes realistic retinal fundus images.
• Data augmentation with synthesized images improved the accuracy by 1.53% and quadratic weighted kappa by 1.70% for a state-of-the-art DR grading model.

Conclusions:

• The proposed RF-GANs method effectively addresses data imbalance in diabetic retinopathy detection.
• Synthesized retinal fundus images enhance the performance of deep learning models for DR grading.
• This approach offers a promising solution for improving the accuracy and efficiency of DR diagnosis.