Forecasting the diabetic retinopathy progression using generative adversarial networks.

Abstract

Background: Diabetic retinopathy (DR) is the leading cause of blindness worldwide, making early prediction of DR progression crucial for effectively preventing visual loss. This study introduces a prediction framework DRForecastGAN (Diabetic Retinopathy Forecast Generative Adversarial Network), and investigates its clinical value in predicting DR development.

Methods: DRForecastGAN model, consisting of a generator, discriminator, and registration network, was trained, validated, and tested in training (12,852 images), internal validation (2734 images), and external test (8523 images) datasets. A pre-trained ResNet50 classification model identified the DR severity on synthetic images. The performance of the proposed DRForecastGAN model was compared with the CycleGAN and Pix2Pix models in image reality and DR severity of the synthesized fundus images by calculating Fréchet Inception Distance (FID), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and area under the curve (AUC).

Results: DRForecastGAN model has the lowest FID, highest PSNR and highest SSIM on internal validation (FID: 27.3 vs. 32.8 vs. 34.4; PSNR: 25.3 vs. 17.0 vs. 16.9; SSIM: 0.93 vs. 0.79 vs. 0.65) and external test (FID: 37.6 vs.45.1 vs.48.4; PSNR: 20.7 vs.15.2 vs.14.7; SSIM: 0.86 vs.0.69 vs.0.63) datasets compared with Pix2Pix and CycleGAN models. In the prediction of DR severity, our DRForecastGAN model outperforms both Pix2Pix and CycleGAN models, achieving the highest AUC values on both internal validation (0.87 vs. 0.76 vs. 0.75) and external test (0.85 vs. 0.70 vs. 0.69) datasets.

Conclusions: The proposed DRForecastGAN model can effectively visualize DR development by synthesizing future fundus images, offering potential utility for both treatment and ongoing monitoring of DR.