Research On Artifact Suppression In Low-dose CT Imaging For Classical Filtered Back-projection Algorithm

Computed tomography(CT)obtains the reconstructed anatomical structure images of the human body based on a non-invasive scanning manner,which are reconstructed with the X-ray projection data measured from the detector.Nevertheless,X-rays induce radiation harm to the human body,and excessive radiation exposure increases the risk of cancer.In the modern CT imaging,the current research focuses on reducing the CT radiation dose as much as possible while providing image quality that meets clinical requirements.Many low-dose CT imaging technologies and application researches have been developed.In low-dose CT imaging technology,low-mAs and sparse-view imaging effectively reduce the radiation dose,but the reconstruction algorithm represented by filtered back-projection(FBP)introduces artifacts in CT images,resulting in low diagnostic efficiency.Deep learning based low-dose CT imaging method has recently outperformed traditional methods in terms of noise reduction and artifact suppression.However,the current algorithms still have certain limitations.First,the deep learning model trained with paired simulated data achieves outstanding performance,while the performance of this model transferred to the real clinical data is decreased.Second,data domain of the current deep learning based methods is usually restricted to the projection domain,image domain,and projection-image combination domain.To address the aforementioned fundamental difficulties,this thesis seeks to conduct the following two investigations for the classical FBP algorithm,which integrates the characteristics of projection data and view-by-view back-projection tensor(VVBPTensor)data in low-dose CT imaging strategy:(1)Aiming at the classical FBP algorithm,we propose a low-dose CT projection denoising model based on self-supervised learning,restoring the noisy projection to high-performance projection.In the event of unlabeled low-dose level projections,the model utilizes the structural similarity of adjacent projections.The model takes multiple consecutive adjacent projections and the middle projection as the input and the label to denoise the middle projection,using the inter-view gradient restriction to enhance detail recovery additionally.Experimental results of simulated low-dose patient data in CT Challenge and real low-dose phantom data show that the proposed model achieves better image quality than other unsupervised learning models.In the experiments of two low-dose levels,this method is validated to be robust and has superior image recovery performance at lower dose level.(2)Using the VVBP-Tensor space and its characteristics in the classical FBP algorithm,we propose a deep learning framework based on sorting VVBP-Tensor for low-dose CT imaging,as well as image segmentation.The sorting VVBP-Tensor data is the intermediate data between the projection and image data,which contains uncompressed reconstruction information that reflects structural similarity and tensor sparsity.In the sparse view scan protocol,the deep learning model based on new data domain achieves artifact suppression and image segmentation.Experimental results of different sparse-view sampling rates demonstrate that the proposed framework provides superior image quality than other compared data domain methods.Moreover,the sorting VVBP-Tensor domain used in this framework extracts organ contours more accurately than the sparse-view image training method.