Comparative Evaluation of Deep Generative Models for Predicting 12-Month Neovascular AMD Progression Using OCT and...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Neovascular age-related macular degeneration (nAMD) requires accurate long-term anatomic progression prediction for effective management.
• Predicting nAMD progression from pretreatment retinal images can aid in personalized treatment planning.

Purpose of the Study:

• To systematically compare six deep generative models for predicting long-term nAMD anatomic progression.
• To evaluate the performance of GAN-based and diffusion-based models using quantitative metrics and expert grading.

Main Methods:

• Retrospective analysis of OCT and fundus images from 85 treatment-naïve nAMD eyes.
• Training of five GANs (BiCycleGAN, CycleGAN, Pix2pixHD, CycleGAN-Turbo, Pix2pix-Turbo) and one diffusion model (Stable Diffusion Img2Img).
• Quantitative assessment using SSIM, PSNR, MSE, RMSE, and a visual Turing test by expert graders.

Main Results:

• Pix2pixHD achieved the highest quantitative image quality metrics (SSIM, PSNR) across models and modalities.
• Experts identified synthetic images in 58% of cases, with OCT images showing near-chance discriminability (52%).
• Pix2pixHD generated the most clinically realistic images, though not reliably distinguishable from real images by experts.

Conclusions:

• Deep generative models, particularly Pix2pixHD, show potential for predicting nAMD progression with high fidelity.
• These models can generate clinically realistic synthetic retinal images, aiding AI-driven decision support in personalized retinal care.
• This technology can shift nAMD management towards proactive paradigms, improving patient counseling and treatment planning.