Reconstruction Of Sparse View CT Images Based On Deep Learning

Computed tomography is one of the most advanced medical imaging technologies.which uses X-rays to reconstruct the attenuation signal obtained when the object passes through the material layer,so as to obtain the tomographic image signal of the object.Although there is no evidence that the low doses of radiation used in CT scans cause long-term harm,the potential risk of cancer may increase slightly with increasing radiation doses.But we still hope to minimize radiation damage to the human body,and obtain clear CT images at the same time.However,due to incomplete projection data in sparse-view CT,there are severe streak artifacts over the most widely used commercially available filtered back projection reconstruction images.In recent years,deep-learning with convolutional neural networks is being developed for image reconstruction from sparse-view projection data.In this paper,the problem of CT image reconstruction from sparse view has been studied in depth.The deep learning-based CT reconstruction algorithm studied in this paper is to reduce the number of samples of projection data and improve the imaging quality and imaging speed of projection sparse-view data on the premise of reducing the CT radiation dose.The physical principles of CT imaging and the idea of image reconstruction are expounded systematically in this paper.The training set consists of4000 cases where each case consists of the truth image,128-view sinogram data,and the corresponding 128-view FBP image.The networks are trained to predict the truth image from the sinogram and FBP data,which mainly includes upsampling and detail restoration.The upsampling stage uses sinusoidal data to generate a preliminary reconstruction result constrained by a deep supervision loss.The detail recovery stage uses the residual dense block to learn residual information,and then uses local residual learning to adaptively fuse the input of the residual dense block and the output of the last residual layer,by adaptively combining the original low-resolution information and the learned information.Reconstruct high-resolution images.The experimental results show that the imaging speed of the deep learning sparse reconstruction framework is faster,the reconstructed images are also significantly improved,the streak artifacts are reduced,and the texture details are restored more realistically,which can achieve a good reconstruction effect.