MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Blind Image Restoration for Low-Dose Lung CT Using Residual Network and Transfer Learning

A Zhong1*, B Li2 , N Luo1 , Y Xu1 , L Zhou1 , X Zhen1 , (1) Department of Biomedical Engineering, Southern Medical University, Guangzhou, Guangdong, (2) Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangzhou

Presentations

(Monday, 7/15/2019) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 1

Purpose: Deep learning has been extensively investigated to remove artifacts in low-dose CT (LDCT). However, most of these denoisers could not perform well in blind image denoising and will cost massive training time. To solve these problems, we proposed a novel convolutional network (CNN) model to blindly restore LDCT image effectively and accurately.

Methods: In order to reduce the heavy training time of CNN and improve denoiser’s efficiency, we utilized residual network and transfer learning techniques in our model. A residual learning architecture is designed to estimate the difference of noise-free image and its original map, and then the denoised CT image can be obtained by subtracting this map from the input of the model. Our network architecture contains 20 layers, including 1 block of Conv+ReLU, 18 blocks of Conv+BN+ReLU and 1 block of Conv. After pre-training the model using 400 natural images with blind Gaussian noise levels, we fine-tuned this model with lung LDCT dataset of 39 patients. The LDCT data were simulated at noise levels of (10:10:100) mAs. The loss function of our model is mean square error and optimized with stochastic gradient descent. This method was implemented on a GPU platform by using the MatConvnet toolbox. We also compared this proposed method with LDCT and BM3D-denoised images qualitatively and quantitatively.

Results: Our model converged fast within only few epochs by initializing the network parameters with transfer learning. Visual inspection can consistently find that our proposed method suppress noisy artifacts with good edge recovery. Quantitative evaluation shows that PSNR and RMSE values of our denoised images are better than those in LDCT and BM3D-denoised images, with higher PSNR (44.73dB>38.73dB>35.23dB) and lower RMSE (1.5114<3.1547<5.0916), respectively.

Conclusion: Our study has demonstrated that this model can reduce the training time and blindly suppress noisy artifacts in LDCT images effectively.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email