Learning-Based X-Ray Image Denoising Utilizing Model-Based Image Simulations

Abstract

X-ray guidance is an integral part of interventional procedures, but the exposure to ionizing radiation poses a non-negligible threat to patients and clinical staff. Unfortunately, a reduction in the X-ray dose results in a lower signal-to-noise ratio, which may impair the quality of X-ray images. To ensure an acceptable image quality while keeping the X-ray dose as low as possible, it is common practice to use denoising techniques. However, at very low dose levels, the application of conventional denoising techniques may lead to undesirable artifacts or oversmoothing. On the other hand, supervised learning techniques have outperformed conventional techniques in producing suitable results, provided aligned pairs of associated high- and low-dose X-ray images are available. Unfortunately, it is neither acceptable nor possible to acquire such image pairs during a clinical intervention. To enable the use of learning-based methods for the denoising of X-ray images, we propose a novel strategy that involves the use of model-based simulations of low-dose X-ray images during the training phase. We utilize a data-driven normalization step that increases the robustness of the proposed approach to varying amounts of signal-dependent noise associated with different X-ray image acquisition protocols. A quantitative and qualitative analysis based on clinical and phantom data shows that the proposed strategy outperforms well-established conventional X-ray image denoising methods. It also indicates that the proposed approach allows for a significant dose reduction without sacrificing important image information.