Sparse-View CT Image Reconstruction Method Based On Deep Learning

Computed Tomography(CT)is a widely used medical imaging technique in clinical diagnosis.It combines a series of X-ray projections taken from different angles of the human body with computer processing to recover slices of the body’s interior,including soft tissues,blood vessels and bones.However,X-rays produce ionizing radiation in the human body,which may induce a series of diseases such as cancer and leukemia.It is very important to study the reduction of radiation injury while obtaining as clear CT images as possible.At present,there are many ways to reduce radiation dose,among which sparse-view CT reconstruction is an effective method to reduce radiation dose.This thesis conducts an in-depth study on sparseview CT reconstruction and the main work is as follows:The research group proposed the LEARN model based on deep learning in the early stage.Compared with traditional algorithms and post-processing algorithms based on deep learning,this algorithm can achieve high-quality reconstruction results in the case of a small amount of data.Although remarkable imaging results had been achieved,only simple convolutional neural networks are used for the learnable regularization items in the image domain.Due to the complexity of regularization items,simple convolutional neural networks cannot effectively perform representation and optimization.Therefore,to solve this problem,this article firstly studies its improvement in the image domain based on the LEARN model,and proposes a sparse-view CT reconstruction network that combines adaptive attention mechanism.On the clinical CT data sets,the improved method in this thesis can effectively retain the details and structural information of images,and improve the quality of the sparse-view CT images.Up to now,most researches on sparse-view CT reconstruction have focused on the reconstruction of the image domain,while there are few studies on how to select the projection data for reconstruction in the sinogram domain.For instance,in most studies,when selecting sinogram domain data,a certain amount of projection data is selected for reconstruction at equal intervals.However,for different parts of the human body,projection data of different angles may have different effects on the reconstruction results.In response to this problem,this thesis innovatively proposes a learning-based angle selection mechanism to study the importance of projection data for different parts of the human body and different angles.For the existing sparse-view CT reconstruction network,partial projections dynamically selected in the full projection cannot be used for reconstruction.Therefore,this thesis proposes a network that can use either dynamically selected projections or fixed-angle projections for reconstruction.Experimental results show that compared with other reconstruction algorithms,the improved reconstruction network in this article can improve the quality of reconstructed CT images.For different parts of the human body,the angle selection mechanism based on learning can select projection data of different angles for reconstruction.For the same part of the human body,projection data with roughly the same projection angle will be selected for reconstruction,and this article has verified that using the selected projection data for CT reconstruction can obtain better image quality in our experiments.