Removing ring artifacts in CBCT images via generative adversarial networks with unidirectional relative total variation loss

Cone beam computed tomography (CBCT) is an important tool for clinical diagnosis and many industrial applications. However, ring artifacts usually appear in CBCT images, due to device responding inconsistence. This paper designs a generative adversarial network (GAN) to remove ring artifacts and meanwhile to retain important texture details in CBCT images. This method firstly transforms ring artifacts in Cartesian coordinates to stripe artifacts in polar coordinates, which is very helpful for removing ring artifacts. Then, we design a new loss function for GAN, including three parts: unidirectional relative total variation loss, perceptual loss and adversarial loss. Further, inspired by super-resolution generative adversarial networks, we use very deep residual networks for both generator and discriminator. Experimental results show that the proposed method is more effective for ring artifacts removal, compared to our baseline and some traditional methods.