Gradient extraction based multiscale dense cross network for LDCT denoising

Abstract

Low-dose computed tomography (LDCT) reduces the damage caused by ionizing radiation by reducing the X-ray dose. However, CT images reconstructed from low-dose X-rays contain a great deal of noise and artifacts that affect a physician’s diagnosis. We propose a gradient extraction based multiscale dense cross network (GE-MDCNet) for LDCT image denoising. This method adds the gradient images (high-frequency information) obtained from the LDCT image through the gradient extraction network and the shallow feature images (low-frequency information) extracted through the shallow feature extraction block to obtain rich feature maps. The feature images are fed into BackboneNet to obtain better-quality prediction images. In addition, we propose a compound loss function based on the Charbonnier loss and gradient loss to enhance the texture details and improve the visual quality of the images. Extensive experiments on the Mayo and Piglet datasets show that GE-MDCNet can effectively remove noise and artifacts from LDCT images while preserving the image structure and edge information. Compared with existing LDCT image post-processing methods, GE-MDCNet significantly improved all metrics. The algorithm offers a novel concept and method for LDCT imaging and other medical image denoising, which can effectively optimize the imaging quality. Moreover, it improves the accuracy and precision of image diagnosis, which has a profound impact on medical imaging.