Unsupervised Low-dose Cerebral Perfusion CT Image Denoising Based On Attention Mechanis

Acute brain diseases such as acute stroke and transient cerebral ischemia have high morbidity and mortality rates,accounting for 9% of the total death toll each year.Cerebral perfusion computed tomography(CT),as a medical imaging technique for diagnosing brain diseases,can accurately detect the location of brain lesions in a timely manner to save patients’ lives,and plays an indispensable role in clinical practice.However,repeated brain perfusion CT scans will accumulate high radiation doses and cause potential cancer risks to patients,which has aroused widespread concern in the society.Low-dose CT(LDCT)scanning technology can effectively reduce the radiation dose of cerebral perfusion CT scanning,but it will cause serious noise artifacts in the imaging,which greatly affects the clinical diagnosis of doctors.Therefore,it is of great theoretical significance and practical clinical value to carry out LDCT image denoising research to generate cerebral perfusion CT images that meet the diagnostic needs of clinicians.In recent years,LDCT image denoising methods based on deep neural networks have shown good development potential.Current methods mostly use supervised learning strategies,which require a large number of paired training samples consisting of normal-dose CT(NDCT)and LDCT images of patients at the same slice position.However,clinically,due to the ethical constraints of radiation dose,there is no condition for repeated CT scans on patients,so it is extremely difficult to obtain paired training samples,which affects its wider clinical application to a certain extent.The LDCT denoising method based on unsupervised deep neural network learning can realize the denoising function without pairing samples,and has a good application prospect in clinic.In this paper,according to the imaging characteristics of low-dose cerebral perfusion CT and on the basis of the Noise2 Noise unsupervised denoising theory,we carried out the research on unsupervised low-dose cerebral perfusion CT image denoising network based on the self-attention mechanism to improve the imaging quality of low-dose cerebral perfusion CT.The main work of this thesis includes the following two parts:(1)We propose an unsupervised denoising method for low-dose brain perfusion CT images based on deep residual global contextual attention network.For the training samples,in order to make full use of the relevant information between brain CT slices,we adopt three-dimensional CT volume scans as the training samples.To meet the sample demands of the Noise2 Noise unsupervised learning strategy,the current low-dose perfusion CT volume scans is used as the input sample,the average of two adjacent time frame volume scans is used as the target sample.As for the network structure,this thesis first proposes a 3D global context(3D-GC)attention module to describe the correlation of non-local correlation of similar tissue structures in CT images,and then proposes a deep residual global context network(RGCN)model based on 3D-GC.The proposed network is composed of multiple residual global context blocks(RGCB),and the modules are connected by a residual path,so as to better extract the deep semantic information of the image and improve the image detail generation ability of the network.The simulation and real data experiment results show that the proposed method can achieve better artifact noise suppression effect in low-dose cerebral perfusion CT images than traditional methods.(2)We propose an unsupervised low-dose cerebral perfusion CT image denoising network fully based on 3D context Transformers.In order to better describe the similarity of the same tissue structures in 3D brain perfusion CT volume scans and the statistical distribution differences between different tissues,this thesis adopts the 3D context Transformer(3D-Co T)attention module as the basic module,and proposes an residual encoding and decoding network in the Noise2 Noise framework which fully bases on the 3D-Co T module instead of the convolution operations,thus forming a full attention unsupervised denoising network.The proposed network can effectively overcome the shortcomings of the traditional convolution-based network,such as the limited receptive field and the weak ability to express dynamic context information,and thus greatly improve the representation ability of the three-dimensional contextual semantic information in the brain perfusion CT scans.The experimental results of simulation and real data show that the proposed network can better preserve the detailed structural information while effectively suppress the artifact noise in the low-dose cerebral perfusion CT image than the traditional method.