Research On Model-Guided Deep Learning Single-View Computed Tomography

X-ray computed tomography(CT)can clearly present the 3D structure of the target object,and thus is widely applied in clinical diagnosis and industrial nondestructive testing.However,there are certain limitations in CT reconstruction due to the need of collecting hundreds of projections,which is time-consuming and results in high-dose radiation.In order to solve the above problem,previous studies try to reduce the number of projections and propose CT reconstruction methods based on limited-view,sparse-view,or even single-view projection.Although the rapid development of deep learning has accelerated the research on single-view CT reconstruction in recent years,most of the current methods are completely data-driven and have poor interpretability.In this thesis,a model-guided deep learning single-view CT reconstruction method is proposed by combining the traditional CT imaging model with the powerful deep learning network.The main research contents are as follows:In this thesis,a single-view CT reconstruction framework is proposed,which includes a projection completion network,Filtered Back Projection(FBP)module,and CT fine-tuning network.The FBP module provides model guidance for deep learning with the help of analytical relations in traditional CT models.Different from other purely data-driven single-view reconstruction methods,the proposed method improves the interpretability of CT reconstruction.Since the CT model requires sufficient projection data to complete analytical reconstruction,this thesis proposes a projection completion network before the FBP module,which predicts full-angle projections according to the input single-view projection.However,due to the lack of spatial information,there are certain errors in the predicted projections,which are back-projected into the CT images by the FBP module,resulting in artifacts.To remove artifacts in CT images reconstructed by the FBP module,this thesis proposes a CT fine-tuning network based on adversarial training to provide more accurate CT images.In this thesis,the 4D CT datasets containing 20 lung cancer cases are built from the open-source TCIA database,and the comparisons with other popular single-view CT methods prove that the model-guided deep learning method in this thesis has achieved the current best reconstruction performance.Besides,this thesis builds two purely data-driven frameworks with similar parameter numbers which verify the effectiveness of the model-guided framework in single-view CT reconstruction.In ablation experiments,the gradient image input of projection and generative adversarial training in this thesis proved effective in improving reconstruction quality.In addition,this thesis adopts proper data processing method and training strategy which make the network training more stable.