Low-Dose CT Image Denoising Using a Generative Adversarial Network With a Hybrid Loss Function for Noise Learning

Abstract

Potential risk of X-ray radiation from computed tomography (CT) has been a concern of the public. However, simply decreasing the dose will degrade quality of the CT images and compromise diagnostic performance. In this paper, we propose a noise learning generative adversarial network coupling with least squares, structural similarity and L1 losses for low-dose CT denoising. In our method, noise distributed in the input low-dose CT image is learned by the generator network and then subtracted from the input to generate the final denoised version. The denoised CT images are penalized by the least squares loss function, and they are pulled toward boundary of the decision even though they are classified as normal-dose CT. Least squares stabilize the training process without regularization. Structural similarity and L1 losses are utilized to keep textural details and sharpness of the denoised CT images respectively. Experiments and results show that our method can effectively suppress noise and remove artifacts compared with the state-ofthe-art methods. The texture statistical properties, which include mean, standard deviation, uniformity, and entropy, further confirm that the generated noise-reduced CT image is as close as to that of the normal-dose counterpart. INDEX TERMS Deep learning, generative adversarial network, least squares, low-dose CT, denoising.

Low‐dose CT image and projection dataset

“…At the simulated reduced dose levels included in this library, we believe that any differences that may exist between simulated and measured data have negligible impact on algorithms developed using the provided lower‐dose data. This belief is supported by the successful use of the data in numerous publications 26–34 …”

“…A residual encoder–decoder convolutional neural network (RED‐CNN) for 2D and 3D CT denoising 26,27 A generative adversarial network (GAN) for low‐dose CT denoising 28,29 A multiresolution deep learning U‐net for sparse‐view CT 30 …”

“…The dataset is provided in a paired fashion at both routine dose and simulated lower dose, which is ideal for many techniques involving supervised learning. Many noise reduction techniques trained based on these paired data have demonstrated great success in terms of reducing image noise and improving image quality 26–34 . However, given the recent success of unsupervised low‐dose CT reconstruction, it is desirable also to have unpaired low‐dose and full‐dose data available, which is a topic of interest for future development.…”

“…This belief is supported by the successful use of the data in numerous publications. [26][27][28][29][30][31][32][33][34] The dataset is provided in a paired fashion at both routine dose and simulated lower dose, which is ideal for many techniques involving supervised learning. Many noise reduction techniques trained based on these paired datasets have demonstrated great success in terms of reducing image noise and improving image quality.…”

“…Many noise reduction techniques trained based on these paired datasets have demonstrated great success in terms of reducing image noise and improving image quality. [26][27][28][29][30][31][32][33][34] However, given the recent success of unsupervised low-dose CT reconstruction, it is desirable also to have unpaired low-dose and full-dose dataset available, which is a topic of interest for future development. Meanwhile, the current paired dataset provided in this work can still be used to meet that purpose using appropriate arrangement of training cases (e.g., organize the cases in an unpaired fashion), which is beneficial as the full-dose images can be used as references even if they are not used in the training and testing.…”