Research On Low-dose CT Image Denoising Based On Full Attention Mechanis

X-ray computed tomography(CT)has been widely used in the clinical diagnosis ofvarious diseases and has become an indispensable auxiliary diagnosti c tool in clinical medicine.However,the ionizing radiation generated by X-ray in the process of CT imaging has a potential carcinogenic risk to human body,which has caused widespread global concern.Low-Dose CT(LDCT)can effectively reduce the radiation dose applied to patients during CT imaging.On the other hand,scanning under low-dose conditions will cause serious noise and artifacts in CT images,resulting in degradation of image quality,which is difficult to meet the needs of clinical diagnosis.It is of great theoretical significance and practical value to suppress the noise of LDCT image to obtain clinical satisfactory image quality.Recently the deep neural network method has made important progress in the field of LDCT image denoising.However,the current network model is mostly based on convolution operation,and its inherent characteristics of convolution kernel weight sharing and receptive field limitation affect the network’s ability to exprexpress local region features of images with different distribution properties to a certain extent.For this reason,this thesis introduces the context trasformer(CoT)attention mechanism,and constructs a LDCT image denoising network model based on full attention.The specific work of this thesis is as follows:(1)A full attention low-dose CT image denoising network based on context transformer(CoT)is proposed.The proposed network adopts the framework of generating adversarial network(GAN),and uses CoT module to completely replace the convolution structure in the traditio-nal network to form a full attention network model.Specifically,both the generator and discri-minator use CoT as the basic module,forming an end-to-end network architecture of full Tran-sformer.The proposed network model can effectively solve the problem of limited receptive fi-eld of traditional convolutional network,and the static context and dynamic context structure in CoT module can better represent the local correlation of local areas in CT images,as well as the structural changes of different tissues and organs.The visual evaluation and quantitative evalua-tion results show that the proposed network can achieve better effect of suppressing artifacts in LDCT images than the traditional network model.(2)A low-dose CT image denoising algorithm based on multi-scale contextual transformer(MsCoT)residual encoding and decoding is proposed.In order to better express the scale change characteristics of different tissues and organs in CT images,this thesis proposes a multiscale context transformer(MsCoT)module.Inspired by Multi Res UNet,the proposed MsCoT uses three lightweight 3×3 convolution sequences to replace the fixed-size convolution operators in the original CoT,and forms multi-scale static context features that can simultaneously express 3×3,5×5,and 7×7 convolutions by splicing the different outputs of the convolution sequence,and learns dynamic context under its guidance to obtain multi-scale dynamic context representation.Based on MsCoT,this thesis proposes a new residual in residual encoder-decoder(RIRED)generation network architecture to better recover the details and textures in CT images.The visual and quantitative experimental results show that the proposed network can effectively remove the artifacts in the LDCT image while maintaining the texture details in the CT image better than the traditional network model.