Coded Aperture Computed Tomography Via U-shape Generative Adversarial Network

The X-ray computed tomography(CT)technology is becoming increasingly important in clinical diagnosis as an effective diagnostic tool in medical imaging.CT scans can provide clear tomographic images of various tissues in the human body,providing doctors with effective diagnostic and treatment plans.However,there is a trade-off between the quality of the reconstruction and the radiation dose caused by a complete CT scan.Traditional analytical and algebraic reconstruction algorithms cannot maintain the quality of reconstructed images while reducing radiation dose.The encoding-aware CT imaging technology combines coding aperture imaging with compressive sensing signal processing to analyze X-ray CT imaging strategies under under-sampled conditions,providing a new method to resolve the contradiction between reconstruction quality and X-ray radiation dose.However,this unsupervised sampling strategy is incompatible with current mainstream CT reconstruction frameworks,and the optimization and reconstruction complexity are high,making it difficult to meet clinical requirements.Therefore,the estimation of missing data in non-continuous CT projections under an unsupervised sampling strategy to make it compatible with analytic CT reconstruction frameworks and to reconstruct clear and clinically relevant CT images has significant research value.To address the limitations of algebraic iterative reconstruction algorithms for nonunderdetermined sampling in encoding aperture CT imaging,this paper proposes a dynamic encoding aperture CT reconstruction algorithm based on a U-Net generative adversarial network to improve the reconstruction quality of coding mask board random spatial X-ray projection missing data.The algorithm constructs a dynamic game model of non-continuous sparse projection based on a U-Net generative adversarial network,including a generator and discriminator.The encoder-decoder structure of the generator preserves more dimension and position information through concatenation to achieve feature fusion,mainly predicting random projection data of the encoding aperture.The discriminator uses pixel-wise feedback to regularize image consistency to distinguish whether the output data is real or fake.Meanwhile,the joint loss function composed of adversarial loss,reconstruction loss,and content loss is used to constrain the output results of the network to achieve better prediction results.The trained generator predicts the newly acquired encoding aperture projection data and uses the FBP reconstruction algorithm to reconstruct the original CT image from the two-dimensional projection data with the encoding aperture predicted in the training stage.A series of training and testing results show that the dynamic encoding aperture CT imaging can achieve efficient analytic(non-iterative)reconstruction under non-continuous sam pling conditions by combining a fast and low-memory encoding aperture optimization method.To improve the reconstruction quality of encoding mask board in CT images with specific spatial X-ray projection missing data,this paper proposes a static encoding aperture CT reconstruction algorithm based on a deep convolutional generative adversarial network.The algorithm mainly adds a perceptual loss function to the joint loss function of the dynamic reconstruction algorithm,as well as feature extraction of projection data in specific spaces.Using the pre-trained VGG16 network as the fixed network for perceptual loss,real images and predicted images are used as inputs to obtain the corresponding output features,and then the L2 loss is constructed based on the output features.Compared with the ordinary L2 loss,the output feature details are enhanced by approximating the perceptual information between the real image and the network-generated result.To solve the problem of image quality degradation caused by metal artifacts in CT reconstruction,this paper proposes an end-to-end generative adversarial network-based algorithm to remove metal artifacts in CT images.During the training phase,the metal artifact projection images with sineshaped beams are used as inputs for the generative network,which predicts the parts of the projection image masked by the sine-shaped metal artifact.During the testing phase,new metal artifact projection images with sine-shaped beams are input into the trained generative network,which predicts the metal artifact image and outputs the complete projection data.Finally,FBP reconstruction is performed on the complete projection data to obtain CT images without metal artifacts.Experimental results show that this algorithm can effectively improve CT image quality in most 2D projection data with metal artifacts.