Research On Low-dose CT Image Denoising Based On Unsupervised Attention Networ

Computed tomography(CT)has been widely used in clinical diagnosis because it can obtain high-resolution images of patient’s tissue structures in a non-invasive manner.However,the high X-rays ionizing radiation produced during CT scans has a risk of cancer for patients,which has caused worldwide public concern.In recent years,deep learning-based methods have shown good performance for low-dose CT(LDCT)image denoising.Most existing deep learning-based LDCT image denoising methods adopt supervised learning strategy,which requires a large number of well-paired sample images for network training which consists of pairs noisy images and clean images.However,on the other hand,it is very difficult to obtain well-paired CT training sample images in a clinical setting constrained by radiation control,which limits the clinical application of the deep learning-based methods.Unsupervised learning methods do not require well-paired training samples and have broader clinical applicability compared with supervised learning methods.This thesis aims to address the LDCT image denoising problem in the Cycle GAN unsupervised deep learning framework.By constructing new attention mechanisms and deep denoising network models,we carry out research on low-dose CT image denoising based on unsupervised attention networks to improve the quality of LDCT images.The main works of this thesis are as follows:(1)We propose an unsupervised low-dose CT image denoising network based on a channel selective spatial pyramidal attention mechanism.This thesis introduces the Spatial Pyramid Attention(SPA)mechanism to describe the global contextual information of different scales in CT images.In addition,this thsis further proposes a channel selective SPA(CSSPA)model on the basis of the original SPA structure for better fusing of the original features and attentional features.Within the Cycle GAN framework,we propose an unsupervised LDCT image denoising network based on CSSPA(called CSSPA-Cycle GAN).Quantitative and qualitative experimental results demonstrate that the proposed method can achieve better LDCT image noise suppression performance compared with traditional unsupervised learning methods.(2)We propose an end-to-end unsupervised low-dose CT image denoising network based on the full attention module.Different from the traditional network structures which are based on convolutional operations,this thesis adopts the Contextual Transformer(Co T)attention module as the basic network composition module for better describing the local similarity of similar tissue structures in CT images and the variability of data statistical distribution of different tissues.In addition,on the basis of the Co Tthis thesis further proposes the multi-scale Contextual Transformer(MCo T)module for better representing the contextual semantic information of different fineness in CT images.Within the Cycle GAN framework,this thesis proposes an unsupervised low-dose CT image denoising network fully based on the MCo T module(called MCo T-Cycle GAN).In the proposed MCo T-Cycle GAN network,both the generator and discriminator use the MCo T as the basic module,thus forming an all-Transformer network architecture.Compared with traditional network structures with convolutional operation as the basic module,the proposed network can improve the contextual semantic representation ability for CT images,thus helping to better restore the fine structures in CT images.Quantitative and qualitative experimental results show that the proposed MCo T-Cycle GAN network can effectively preserve the textural details while suppressing the artifact noise for LDCT images.