Scholarlink

<p>Low-dose (LD) PET/CT imaging presents a critical challenge in medical imaging, where reducing radiation exposure while preserving diagnostic quality requires advanced computational approaches. This study presents Enhanced Pix2Pix (Enh-Pix2Pix), an advanced conditional generative adversarial network (cGAN) for low-count PET image quality enhancement, where only the PET acquisition count level is reduced while the CT is acquired at a standard dose for attenuation correction purposes. The framework introduces four synergistic components: Xavier normal initialization for stable early-training gradient flow, selective decoder dropout to prevent noise overfitting while preserving fine-detail reconstruction, gradient norm clipping to bound training instability from extreme noise batches, and LSGAN loss to maintain continuous adversarial feedback throughout training. Together, these modifications collectively address training instability, overfitting, and gradient saturation without increasing model complexity or computational cost. Evaluation across a medical PET/CT dataset under extreme LD conditions (1%, 2%, 5%, and 10% count levels) using fivefold cross-validation (CV) demonstrates superior image quality recovery. The model achieved RMSE values ranging from 0.01 ± 0.017 SUV (standard uptake value) at 10% counts to 0.033 ± 0.05 SUV at 1% counts, with corresponding MAE values between 0.004 ± 0.004 SUV and 0.010 ± 0.0127 SUV, consistently outperforming baseline Pix2Pix by 8.9–32.6% and achieving up to 52.7% RMSE improvements over state-of-the-art methods under extreme LD conditions. The Enh-Pix2Pix framework demonstrates exceptional pixel-level reconstruction fidelity in challenging LD conditions, establishing it as a clinically viable solution for radiation dose reduction while maintaining diagnostic accuracy. This framework further highlights the feasibility of integrating advanced adversarial learning into clinical PET/CT pipelines, paving the way for safer, faster, and more cost-effective imaging protocols. The Python code from our study is publicly accessible at: <a href="https://github.com/jafarMajidpour/Enhanced-Adversarial-Learning-for-Image-quality-Restoration-in-Low-Dose-PET-CT-Imaging">https://github.com/jafarMajidpour/Enhanced-Adversarial-Learning-for-Image-quality-Restoration-in-Low-Dose-PET-CT-Imaging</a></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

Enhanced adversarial learning for image quality restoration in low-dose PET/CT imaging

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Jafar Majidpour,
  • Mohammad Saber Azimi,
  • Mohammed H. Ahmed,
  • Habibollah Dadgar,
  • Xiaotong Hong,
  • Sahar Rezaei,
  • Hossein Arabi

摘要

Low-dose (LD) PET/CT imaging presents a critical challenge in medical imaging, where reducing radiation exposure while preserving diagnostic quality requires advanced computational approaches. This study presents Enhanced Pix2Pix (Enh-Pix2Pix), an advanced conditional generative adversarial network (cGAN) for low-count PET image quality enhancement, where only the PET acquisition count level is reduced while the CT is acquired at a standard dose for attenuation correction purposes. The framework introduces four synergistic components: Xavier normal initialization for stable early-training gradient flow, selective decoder dropout to prevent noise overfitting while preserving fine-detail reconstruction, gradient norm clipping to bound training instability from extreme noise batches, and LSGAN loss to maintain continuous adversarial feedback throughout training. Together, these modifications collectively address training instability, overfitting, and gradient saturation without increasing model complexity or computational cost. Evaluation across a medical PET/CT dataset under extreme LD conditions (1%, 2%, 5%, and 10% count levels) using fivefold cross-validation (CV) demonstrates superior image quality recovery. The model achieved RMSE values ranging from 0.01 ± 0.017 SUV (standard uptake value) at 10% counts to 0.033 ± 0.05 SUV at 1% counts, with corresponding MAE values between 0.004 ± 0.004 SUV and 0.010 ± 0.0127 SUV, consistently outperforming baseline Pix2Pix by 8.9–32.6% and achieving up to 52.7% RMSE improvements over state-of-the-art methods under extreme LD conditions. The Enh-Pix2Pix framework demonstrates exceptional pixel-level reconstruction fidelity in challenging LD conditions, establishing it as a clinically viable solution for radiation dose reduction while maintaining diagnostic accuracy. This framework further highlights the feasibility of integrating advanced adversarial learning into clinical PET/CT pipelines, paving the way for safer, faster, and more cost-effective imaging protocols. The Python code from our study is publicly accessible at: https://github.com/jafarMajidpour/Enhanced-Adversarial-Learning-for-Image-quality-Restoration-in-Low-Dose-PET-CT-Imaging