CT-Scan Denoising Using a Charbonnier Loss Generative Adversarial Network

Abstract

We propose a Generative Adversarial Network (GAN) optimized for noise reduction in CT-scans. The objective of CT scan denoising is to obtain higher quality imagery using a lower radiation exposure to the patient. Recent work in computer vision has shown that the use of Charbonnier distance as a term in the perceptual loss of a GAN can improve the performance of image reconstruction and video super-resolution. However, the use of a Charbonnier structural loss term has not yet been applied or evaluated for the purpose of CT scan denoising. Our proposed GAN makes use of a Wasserstein adversarial loss, a pretrained VGG19 perceptual loss, as well as a Charbonnier distance structural loss. We evaluate our approach using both applied Poisson noise distribution in order to simulate low-dose CT imagery, as well as using an anthropomorphic thoracic phantom at different exposure levels. Our evaluation criteria are Peek Signal to Noise (PSNR) as well as Structured Similarity (SSIM) of the denoised images, and we compare the results of our method versus recent state of the art deep denoising GANs. In addition, we report global noise through uniform soft tissue mediums. Our findings show that the incorporation of the Charbonnier Loss with the VGG-19 network improves the performance of the denoising as measured with the PSNR and SSIM, and that the method greatly reduces soft tissue noise to levels comparable to the NDCT scan.

Keywords

• CT-scan denoising
• machine learning
• computed tomography
• medical diagnostic imaging
• generative adversarial network