Research On Post-processing Algorithms Of 3D Sparse Angle CT Images Based On Multi-scale U-net

Sparse angle CT reconstruction images mainly refer to the use of a smaller number of projections for reconstruction during CT reconstruction,which helps to reduce the radiation dose to the patient.But the effect is that there are serious strip artifacts in the reconstructed image.Existing post-processing methods based on analytic reconstruction can cause boundary blur and the inability to accurately predict strip artifacts when processing such images,and analytical reconstruction itself lacks noise suppression.It is also difficult to solve the problem of long reconstruction time based on iterative reconstruction.Convolutional neural networks can effectively extract features of different levels of images due to their powerful nonlinear modeling capabilities.In recent years,convolutional neural networks have outstanding performance in the field of image processing,providing a new idea for the study of artifacts in removing sparse angle CT images.Therefore,this paper uses convolutional neural networks to study the suppression and elimination of artifact noise in the image reconstructed by sparse angle CT.The purpose is to reduce the CT scan dose by reducing the number of projection angles of CT reconstruction as much as possible,reduce the radiation generated by the CT scan of the patient,improve the imaging quality,and provide effective assistance for medical diagnosis.In general,this subject can be divided into the following two parts:This paper designs a convolutional neural network for strip artifacts to post-process the sparse angle CT image.The algorithm designs the mapping from noise-containing artifact images to the noise artifacts contained therein.The network explores the multi-scale U-shaped processing structure,residual module,training data partition size,and network depth impact on network performance.The final experimental results show that using the U-shaped structure?the resnet basic module and Appropriate increase in depth of the net can improve the network performance.The larger block size can also make the training more stable.If the block is too small,the artifact structure will be destroyed and the model will not converge.After fully considering the three-dimensional correlation of the tissue structure such as blood vessels in CT data,this topic extends the two-dimensional de-artifact network model to three-dimensional.Finally,the experiment proves that the three-dimensional model can better distinguish the details and further enhance the final removal of artifacts and noise.Analytical reconstruction lacks the suppression of noise.Therefore,The model trained by the complete analysis of the reconstruction result as the reference data will also have some noise and artifacts in the final processing result.in view of this,according to the characteristics of noise and artifacts,The total variation prior information of the final result image is regularized and added to the loss function of the model to further suppress artifacts and smooth noise.There are two kinds of regularization terms designed in this subject: L1 and L2 regularization.The final experimental results prove that the model with L1 regularization can effectively further suppress the artifacts and smooth the noise.