面向低剂量CT图像高分辨率重建的深度强化学习框架HRDRL-Net

Key Points

• The research aims to enhance low-dose computed tomography image quality by balancing noise suppression and structural preservation using a deep reinforcement learning framework.
• Modeling the LDCT denoising task as a sequential decision-making process.
• Using an asynchronous advantage actor-critic algorithm for agent training.
• Constructing a dual-path cooperative architecture and a noise suppression module.
• Designing a composite reward function incorporating pixel-level error, gradient similarity, and local variance constraints.
• HRDRL-Net outperforms mainstream baseline methods in quantitative metrics on Mayo and Piglet datasets.
• On the Mayo test set, HRDRL-Net shows an average increase in peak signal-to-noise ratio by approximately 1.3% and a structural similarity index increase of about 0.7%.
• On the Piglet dataset, it achieves average improvements of 1.5% and 0.8% in peak signal-to-noise ratio and structural similarity index, respectively.