Robust Denoising of Low-Dose CT Images using Convolutional Neural Networks

X-ray computed tomography (CT) images are widely used in medical diagnosis. A drawback of X-ray CT imaging is that the X-rays are harmful with high-dose. Reducing the X-ray dose can reduce the risks but introduce noise and artifacts in the reconstructed image. This paper presents a method, called FD-VGG for denoising of low-dose CT images. FD-VGG estimates the normal-dose image from the low-dose image and, hence, reduces noise and artifacts. In FD-VGG the loss function is defined by the combination of the mean square error (MSE) and perception loss. FD-VGG was trained on a dataset of 226200 low-dose and normal dose image pairs from 6 patients and evaluated on 100 low-dose images from 2 other patients. The corresponding normal dose images of these testing low-dose images are considered as standard images for quantitative evaluation. Two metrics namely PSNR and SSIM were used for objective evaluation. The experimental results showed that the proposed FD-VGG network was able to denoise low-dose images efficiently, in comparison with two state-of-the-art methods.