Research On Low-dose CT Image Denoising Based On Spatially Aware Contextual Networ

Computed Tomography（CT）can display image information of organs and tissues in patients in a non-invasive way,and has been widely used in clinical disease diagnosis.However,CT scans are highly radioactive and pose a potential cancer risk to humans.Low-dose CT（Low-dose CT,LDCT）reduces the radiation in the scanning process by controlling the X-ray tube voltage or tube current,but at the same time it also introduces strong streaks and noise to the CT image,resulting in a decline of the CT image quality which affects the doctor’s diagnostic efficiency.Therefore,how to reduce X-ray radiation hazards to human body and improve the imaging quality of LDCT images has become a hot topics in medical image processing research,which has important theoretical significance and practical clinical value.In recent years,with the rapid development of deep learning technology,deep convolutional neural network（CNN）has shown great potential in LDCT image noise suppression.However,the convolution operation in the traditional CNN denoising network has the defect of limited receptive field,which affects the expressive ability of deep contextual semantic features in CT images.To this end,this thesis conducts research on LDCT image denoising based on spatial-aware context network,explores the spatial-aware context attention mechanisms,designs new deep residual network architectures,and constructs loss functions with texture preservation ability to improve the quality of LDCT images,laying the foundation for the clinical application of LDCT.The main work of this thesis is summarized as follows:（1）We propose a LDCT image denoising network based on an adaptive global contextual attention mechanism.In order to describe the correlation of non-local regions and the differences of regional statistical distribution in CT images,this thesis first constructs a lightweight adaptive global context（Adaptive Global Context,AGC）network module to achieve adaptive aggregation in each local neighborhood Purpose of contextual semantic information.On this basis,this thesis further proposes an AGC-based Long-Short Residual Encoder-Decoder Network（AGC-LSRED）to achieve effective denoising of LDCT images.The proposed network consists of Residual Adaptive Global Context Blocks（RAGCB）with skip links to simplify the network training process and better extract structural information in images.Furthermore,this thesis proposes a composite loss that combines L₁ loss,adversarial loss and self-supervised multi-scale perceptual loss to better preserve the fine structure of denoising results.Experimental results show that the proposed method achieves better results in noise suppression and fine structure preservation than traditional methods.（2）We propose a LDCT image denoising network based on a space-aware multi-modal global context attention mechanism.In view of the structural characteristics of CT images,this thesis introduces a spatially aware context attention mechanism（Spatially-Aware Context,SAC）with multi-modal representation capabilities to better describe the multi-modal contextual semantic information formed by various organs and tissues in CT images.Based on SAC,this thesis further constructs the residual SAC module（Residual Spatially Aware Context Block,RSACB）,and introduces it into the residual encoder-decoder network architecture to form an LDCT denoising with spatially aware multi-mode global context expression ability network.For the loss function,on the basis of L₁ loss,adversarial loss and self-supervised multi-scale perception loss,this thesis further introduces the Joint Sparsity Transform（JST）loss to better balance the noise suppression and texture details of the proposed network.Qualitative and quantitative experimental results show that the proposed method is superior to traditional methods in LDCT image noise suppression and texture preservation.