Unsupervised Training of a Dynamic Context-Aware Deep Denoising Framework for Low-Dose Fluoroscopic Imaging

Low-dose fluoroscopy is essential for real-time X-ray visualization, supporting dynamic diagnostic assessments while minimizing harmful radiation exposure to patients. However, low-dose imaging introduces noise that can impair diagnostic accuracy. Although numerous deep learning methods have been developed for noise reduction in medical imaging, the unique challenges of fluoroscopy - such as motion artifacts due to its real-time nature, limited access to clean reference data, and high noise levels - diminish the effectiveness of current deep learning-based denoising techniques, leaving research in this area relatively constrained. To address these challenges, we present three key innovations. First, we propose an unsupervised framework for dynamic, context-aware denoising in fluoroscopy, introducing the multiscale recurrent attention U-Net (MSR2AU-Net) to effectively reduce noise without clean data by directly targeting initial noise. Second, our proposed dual-noise suppression strategy combines a knowledge distillation-based module for uncorrelated noise with a recursive filtering module for correlated noise, enhancing both denoising quality and motion stability. Finally, we design a pixel-wise dynamic object motion cross-fusion matrix combined with an edge-preserving loss function to preserve fine details amidst motion changes. Our model was evaluated on 3500 fluoroscopy images from dynamic phantoms (2400 for training and 1100 for testing) and 350 clinical images from spinal surgery patients, confirming its effectiveness in clinical settings. To further validate the robustness of our approach, we also tested it on the "2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge"dataset, using 4800 images for training and 1136 for testing, demonstrating efficacy across both fluoroscopy and CT imaging. Quantitative evaluations demonstrated that our approach outperformed existing unsupervised algorithms, achieving a structural similarity index measure (SSIM) of 0.9803 and a peak signal-to-noise ratio (PSNR) of 39.12 dB on the dynamic phantom dataset, and an SSIM of 0.9591 with a PSNR of 36.62 dB on the Mayo CT dataset. These results indicate that our approach surpasses state-of-the-art unsupervised algorithms in both visual quality and quantitative metrics, achieving performance comparable to well-established supervised methods in low-dose fluoroscopy and CT imaging. The source code will be available at https://github.com/sunyoungIT/UDCA-Net.git.