Image Restoration for Low-Dose CT via Transfer Learning and Residual Network

Anni Zhong,Ning Luo,Xin Zhen,Yuan Xu,Linghong Zhou,Bin Li

doi:10.1109/access.2020.3002534

Abstract

Deep learning has recently been extensively investigated to remove artifacts in low-dose computed tomography (LDCT). However, the power of transfer learning for medical image denoising tasks has not been fully explored. In this work, we proposed a transfer learning residual convolutional neural network (TLR-CNN) to restore LDCT images at single and blind noise levels. A residual network was implemented to effectively estimate the difference between denoised image and its original map, and a noise-free image was obtained by subtracting the residual map from the LDCT image. The results were compared to competing baseline denoising methods in terms of quantitative metrics including the PSNR, RMSE, SSIM and FSIM. For the single noise level, the proposed method demonstrated better denoising performance than the other algorithms for both simulation data and clinical data. As for the blind denoising, the image qualities were improved for all noise levels for all the quantitative metrics, but such improvements were decreasing as the noise level decrease (higher mAs). Comparative experiments suggested that the proposed network could effectively suppress artifacts and preserve image details with faster converge rate and reduced computational time.

Highlights

X-ray computed tomography (CT) is widely used to help doctors diagnose diseases and guide surgery and radiotherapy through imaging [1], [2]
To cope with the problems associated with low-dose CT (LDCT), a number of algorithms have been designed; these algorithms can be roughly divided into three categories, namely, (a) sinogram filtering techniques, (b) iterative reconstruction, and (c) postprocessing methods
We explored the feasibility of applying transfer learning to a residual neural network (TLR-convolutional neural networks (CNNs)) for LDCT image denoising

Summary

Introduction

X-ray computed tomography (CT) is widely used to help doctors diagnose diseases and guide surgery and radiotherapy through imaging [1], [2]. Representative traditional methods to denoise in sinogram domain mainly includes bilateral filtering [4], structural adaptive filtering [5] and penalized likelihood methods [6], [7] These approaches, suffer from the difficulty of raw data acquisition in clinical practices, as well as the potential loss of image edges over-smoothed by filters. Used image priors includes total variation [8]–[10], nonlocal means (NLM) [11], [12], dictionary learning [13], tight wavelet frames [14], and low-rank matrix decomposition [15] These methods require excessive computational times because of the back-and-forth iteration between the projection and

Methods

Results

Conclusion