Denoising Chest LDCT Images Using a Variant Unet Based on Hybrid Attention Mechanism and Diffusion Model

Background: Computed Tomography (CT) may increase cancer risk due to high-dose radiation exposure. Although Low-Dose CT (LDCT) reduces radiation, its noise compromises diagnostic accuracy. This study deeply optimized LDCT image denoising techniques to balance image quality and noise reduction effects. Methods: This study investigates three LDCT image denoising methods combining Unet and diffusion models. The UDiff model enhances image quality by adding Gaussian noise and employing Unet for denoising. The UAdiff model omits the noise-adding step and introduces a hybrid attention mechanism to strengthen feature extraction. The DUPAnet model utilizes LDCT images as prior information, integrating a hybrid attention mechanism for efficient denoising. Experiments were conducted using chest phantom data acquired from multi-vendor CT scanners, optimizing the models with a hybrid loss function, and evaluating denoising performance through various metrics during 2024. Results: The UDiff model exhibits denoising blurring and poor stability when processing similar phantom data. The UAdiff model weakens attention feature capture, leading to detail loss and performance fluctuations. The DUPAnet model shows the best performance in both denoising and detail preservation, with the best performance in both objective metrics Peak Signal-to-Noise Ratio (PSNR) and SSIM (30.37 and 0.87) and subjective images. It enhances the denoising ability of LDCT images, has strong adaptability to various scanning parameters, and has excellent generalization performance. Conclusion: The improved Unet model integrates LDCT priors and a hybrid attention mechanism, effectively denoising while preserving chest LDCT image details, thereby enhancing image quality and generalization capability. Future work will focus on optimizing the algorithm's generalization performance, exploring multimodal fusion and real-time processing to advance clinical applications.