Deep Denoising Network for X-Ray Fluoroscopic Image Sequences of Moving Objects

Prolonged fluoroscopy procedures may involve high patient radiation doses, and a low-dose fluoroscopy protocol has been proven to be effective in reducing doses in an interventional suite. However, the low-dose protocol-caused noise degrades fluoroscopic image quality and then impacts clinical diagnosis accuracy. Here, we propose a novel deep denoising network for low-dose fluoroscopic image sequences of moving objects. The existing deep learning-based denoising approaches showed promising performance in denoising static fluoroscopic images, but their dynamic image denoising performance is relatively poor because they are not able to accurately track moving objects, losing detailed textures of the dynamic objects. To overcome the limitations of current methods, we introduce a self-attention-based network with the incorporation of flow-guided feature parallel warping. Parallel warping is able to jointly extract, align, and propagate features of dynamic objects in adjacent fluoroscopic frames, and self-attention effectively learns long-range spatiotemporal features between the adjacent frames. Our extensive experiments on real datasets of clinically relevant dynamic phantoms reveals that the performance of the proposed method achieves superior performance, both quantitatively and qualitatively, over state-of-the-art methods on a denoising task.