Forecasting the diabetic retinopathy progression using generative adversarial networks.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of global blindness.
• Early prediction of DR progression is critical for preventing vision loss.

Purpose of the Study:

• To introduce and investigate the clinical value of the DRForecastGAN model for predicting DR development.
• To assess the capability of DRForecastGAN in synthesizing future fundus images.

Main Methods:

• The DRForecastGAN model, comprising a generator, discriminator, and registration network, was trained on large datasets.
• Performance was evaluated against CycleGAN and Pix2Pix using metrics like FID, PSNR, SSIM, and AUC.
• A ResNet50 model identified DR severity in synthetic images.

Main Results:

• DRForecastGAN demonstrated superior image quality, evidenced by lower FID and higher PSNR/SSIM compared to other models.
• The model achieved higher AUC values in predicting DR severity on both internal and external validation datasets.
• DRForecastGAN outperformed Pix2Pix and CycleGAN in synthesizing realistic and informative fundus images.

Conclusions:

• DRForecastGAN effectively visualizes DR progression through synthesized future fundus images.
• The model shows potential utility in clinical treatment planning and ongoing patient monitoring for DR.
• This AI-driven approach offers a promising tool for managing diabetic retinopathy.