Unpaired Image Denoising via Wasserstein GAN in Low-Dose CT Image with Multi-Perceptual Loss and Fidelity Loss

Abstract

The use of low-dose computed tomography (LDCT) in medical practice can effectively reduce the radiation risk of patients, but it may increase noise and artefacts, which can compromise diagnostic information. The methods based on deep learning can effectively improve image quality, but most of them use a training set of aligned image pairs, which are difficult to obtain in practice. In order to solve this problem, on the basis of the Wasserstein generative adversarial network (GAN) framework, we propose a generative adversarial network combining multi-perceptual loss and fidelity loss. Multi-perceptual loss uses the high-level semantic features of the image to achieve the purpose of noise suppression by minimizing the difference between the LDCT image and the normal-dose computed tomography (NDCT) image in the feature space. In addition, L2 loss is used to calculate the loss between the generated image and the original image to constrain the difference between the denoised image and the original image, so as to ensure that the image generated by the network using the unpaired images is not distorted. Experiments show that the proposed method performs comparably to the current deep learning methods which utilize paired image for image denoising.

InNetGAN: Inception Network-Based Generative Adversarial Network for Denoising Low-Dose Computed Tomography

“…Apart from that, Li et al [ 37 ] have proposed a WGAN-based self-attention GAN to overcome the limitations of CNN-based LDCT denoising methods. In addition to these applications, recently Yin et al [ 49 ] have proposed a WGAN model with unpaired CT data. They have implemented a multiperceptual loss to determine the feature distribution between the LDCT and RDCT images.…”

Improving Human Image Synthesis with Residual Fast Fourier Transformation and Wasserstein Distance

“…The L style measures the statistical difference between the image generated after the activation layer and the target image. The Wasserstein distance [33] is to solve the problem that GAN is difficult to train and difficult to converge. The advantage of using the W[p, q] to measure the difference between the generated distribution p and the real distribution q is that when there is no intersection or a small intersection between p and q.…”

Dose reduction and image enhancement in micro‐CT using deep learning

“…Wolterink et al 24 were the first ones to propose a generative adversarial network (GAN) for denoising of low dose CT images. Based on the adversarial learning method and the cost function (usually multi-objective functions) different variants exist with Wasserstein GAN and cycle-GAN being the two most studied models for image denoising of low dose CT. [25][26][27][28] For clinical low dose CT, a multitude of DL denoising models have been proposed, and each investigates different combinations of network architectures and loss functions. [14][15][16] Note that, besides image-to-image DL frameworks that serve as post-reconstruction tool, alternatives are to train the DL denoising model in the projection space or to develop an end-to-end DL algorithm that directly maps a noisy sinogram to a clean image (DL-based image reconstruction).…”