Research On Low-dose CT Reconstruction Method Based On Convolutional Neural Network

X-ray computed tomography(CT)is widely used in clinical practice because of its ability to visualize internal organs,bones,soft tissues and blood vessels.In order to obtain high quality CT images,patients are exposed to high doses of radiation during CT imaging,posing a potential risk to their health.Low-dose CT refers to methods of reducing the radiation exposure of patients during CT examinations,including sparse-view CT or reducing tube currents.Both methods result in inadequate reconstruction of highquality images from the original sinogram.Therefore,it becomes increasingly important to recover images as accurately as possible while maintaining a relatively low radiation dose.The recent rapid development of convolutional neural networks has promoted their application in medical imaging.Therefore,this thesis carries out a study of sparse-view CT reconstruction based on deep learning.Since the potential of all existing low-dose CT reconstruction networks for clinical applications is limited by their computation burden,a lightweight sparse-view reconstruction method is first proposed.Meanwhile,inspired by the application of Transformer to natural language processing problems,this thesis introduces Transformer into projection-domain processing networks and proposes an endto-end low-dose CT reconstruction network based on it.The specific research work of this thesis is as follows.(1)Aiming at the limited computing power of reconstruction equipment in the clinical environment,this thesis uses the data features of sparse-view low-dose CT and the advantages of CNN in pixel interpolation to propose a lightweight low-dose CT reconstruction network.The proposed network can be trained to handle low-dose CT projection data with different sparsity levels by modifying the number of output channels of the last convolutional layer.The proposed network can not only reconstruct low-dose CT by connecting with the classical filtered inverse projection module,but also can be embedded into the existing dual-domain reconstruction structure as a projection domain processing module.Experimental results show that the proposed model has high computational efficiency,can be trained on small-scale training datasets,and achieves good reconstruction results at different sparsity levels,demonstrating its potential in clinical applications.(2)To accurately reconstruct CT images,this thesis combines a lightweight convolutional interpolation network and a multi-head attention mechanism to propose a new projection domain sparse angle CT reconstruction method In order to effectively utilize the dual-domain information and perform complementary fusion,this thesis improves the existing dual-domain network architecture and proposes a dual-domain fusion module based on the multi-headed attention mechanism.The improved parallel network structure can effectively avoid the accumulation of errors in the projection-domain data to the image-domain and reduce the coupling between the projection domain module and the image domain module.The fusion module combines the dual-domain information to further refine the reconstructed images and achieve accurate reconstruction.The experimental results show that the proposed model is able to preserve the edge details of CT images at multiple sparsity levels and effectively suppress the noise.