The use of low-dose x-ray fluoroscopy imaging has been found to be effective in reducing radiation exposure during prolonged fluoroscopy procedures that may result in high radiation doses in patients. However, the noise generated by the low-dose protocol can degrade the quality of fluoroscopic images and impact clinical diagnostic accuracy. This paper proposes a novel framework for a low-dose fluoroscopic x-ray denoising algorithm that can recover extremely small details of texture and edges in denoised images. While the existing deep learning–based denoising approaches have shown promising performance, they still exhibit limitations in capturing detailed textures and edges of objects. To address these limitations, we introduce a two-step training framework for denoising. The first network uses multi-frame inputs to leverage more information from several frames, while the second network learns the residual relationship, which can enhance performance in recovering details of texture and edges that the first network may miss. Our extensive experiments on clinically relevant phantoms with real noise demonstrate that the proposed method outperforms state-of-the-art methods in capturing detailed textures and edges in denoised images.
|