Scholarlink

<p>The purpose of this study is to systematically compare six deep generative models for predicting long-term anatomic progression of neovascular age-related macular degeneration (nAMD) from pretreatment retinal imaging. We retrospectively analyzed OCT and fundus images from 85 treatment-naïve eyes initiating anti-VEGF therapy for nAMD. Five GAN-based architectures (BiCycleGAN, CycleGAN, Pix2pixHD, CycleGAN-Turbo, Pix2pix-Turbo) and one diffusion-based model (Stable Diffusion Img2Img) were trained separately for each modality to generate synthetic projections at 3, 6, and 12&#xa0;months. Quantitative performance was assessed using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean squared error (MSE), and root mean squared error (RMSE). Clinical realism was evaluated through a visual Turing test by five blinded expert graders. Pix2pixHD consistently achieved the highest image-quality metrics across all models, modalities, and time points. For OCT, they are SSIM 0.84 and PSNR 27.2&#xa0;dB (3&#xa0;months) and SSIM 0.83 and PSNR 25.8&#xa0;dB (12&#xa0;months). For fundus photographs, they are SSIM 0.80 and PSNR 26.0&#xa0;dB (3&#xa0;months) and SSIM 0.80 and PSNR 24.7&#xa0;dB (12&#xa0;months). In the visual Turing test, experts correctly identified synthetic images in 58% of cases (chance level, 50%), with OCT images showing near-chance discriminability (52%) compared to fundus photographs (64%). This study provides the first systematic comparison of multiple generative architectures for long-term nAMD progression prediction. Pix2pixHD achieved the highest fidelity, generating synthetic images whose realism was frequently, though not reliably, distinguished by observers (58%, not significantly different from chance; <i>p</i> = 0.13), particularly for OCT. These findings support the potential of deep generative models for AI-driven decision support in personalized retinal care. This work bridges the gap between computational science and clinical ophthalmology by demonstrating that deep generative models can transform pretreatment retinal images into clinically realistic predictions of disease progression. By enabling visualization of anticipated anatomical outcomes before treatment initiation, these tools have the potential to transition nAMD management from reactive to proactive paradigms, supporting individualized patient counseling, risk stratification, and evidence-based treatment planning at the point of care.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Fatma Sumer,
  • Murat Toren,
  • Berkutay Asan,
  • Merve Solak,
  • Nurullah Coskuner,
  • Berna Ozkan,
  • V. Levent Karabas,
  • Sibel Demirel,
  • Hakan Ozdemir

摘要

The purpose of this study is to systematically compare six deep generative models for predicting long-term anatomic progression of neovascular age-related macular degeneration (nAMD) from pretreatment retinal imaging. We retrospectively analyzed OCT and fundus images from 85 treatment-naïve eyes initiating anti-VEGF therapy for nAMD. Five GAN-based architectures (BiCycleGAN, CycleGAN, Pix2pixHD, CycleGAN-Turbo, Pix2pix-Turbo) and one diffusion-based model (Stable Diffusion Img2Img) were trained separately for each modality to generate synthetic projections at 3, 6, and 12 months. Quantitative performance was assessed using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean squared error (MSE), and root mean squared error (RMSE). Clinical realism was evaluated through a visual Turing test by five blinded expert graders. Pix2pixHD consistently achieved the highest image-quality metrics across all models, modalities, and time points. For OCT, they are SSIM 0.84 and PSNR 27.2 dB (3 months) and SSIM 0.83 and PSNR 25.8 dB (12 months). For fundus photographs, they are SSIM 0.80 and PSNR 26.0 dB (3 months) and SSIM 0.80 and PSNR 24.7 dB (12 months). In the visual Turing test, experts correctly identified synthetic images in 58% of cases (chance level, 50%), with OCT images showing near-chance discriminability (52%) compared to fundus photographs (64%). This study provides the first systematic comparison of multiple generative architectures for long-term nAMD progression prediction. Pix2pixHD achieved the highest fidelity, generating synthetic images whose realism was frequently, though not reliably, distinguished by observers (58%, not significantly different from chance; p = 0.13), particularly for OCT. These findings support the potential of deep generative models for AI-driven decision support in personalized retinal care. This work bridges the gap between computational science and clinical ophthalmology by demonstrating that deep generative models can transform pretreatment retinal images into clinically realistic predictions of disease progression. By enabling visualization of anticipated anatomical outcomes before treatment initiation, these tools have the potential to transition nAMD management from reactive to proactive paradigms, supporting individualized patient counseling, risk stratification, and evidence-based treatment planning at the point of care.