X-ray Image Blind Denoising in Hybrid Noise Based on Convolutional Neural Networks

Abstract

Low-dose X-ray imaging is a medical imaging method used for disease screening and diagnosis. However, the interpretation of such images is a challenging task because of machine noise. Although some deep learning-based denoising algorithms have made considerable progress, they do not perform well on real X-ray images. Because the actual noise of the X-ray image is more complicated. In this paper, we design a noise model according to the physical principle of X-ray imaging, which is used to simulate the real X-ray image. On this basis, we propose a blind denoising convolutional neural network (X-BDCNN) for low-dose X-ray image enhancement. X-BDCNN consists of two networks. One is used to estimate the noise level of the input noise X-ray image. The other is used to obtain the residual noise image by taking the noisy X-ray image and the estimated noise level as input. The final denoised X-ray image is obtained by subtracting the residual noise image from the input noise X-ray image. In addition, we add a structural similarity (SSIM) loss function to X-BDCNN to maintain the structural information. The experimental results show that the denoising performance of X-BDCNN is better than the existing denoising methods. Code is available online.