Research On Artifact Correction Method Of Human Body Limited Angle Tomography Reconstruction

Computed Tomography(CT)has become one of the indispensable technologies in medical diagnosis.However,due to the limitation of radiation dose and scanning space,sometimes only limited angle projection data can be collected.It leads to seriously artifacts and will reduce image quality,which further affecting the accuracy of diagnosis.In order to solve this problem,many reconstruction methods are proposed,such as projection data completion,iterative reconstruction,deep learning and so on.Among them,deep learning methods have received the most extensive attention,due to the good learning ability,fast processing speed and wide application range.Based on the existing deep learning research,a cascaded residual en-decoder network called Cas-PRED is proposed.The network is a progressive processing method using prior information.Firstly,rough structure is recovered through a residual en-decoder network such as contour and edge.Then the prior information from full angle CT images is integrated into the second network to recover details and improve visual quality.The experiment results show Cas-PRED network can effectively suppress artifacts in limited angle images.Further,due to the limitation of prior information extraction in Cas-PRED network,we propose a Multiloss-RED network based on multiscale loss function,which is more applicable.By introducing feature error in the encoding and decoding process into loss function,the network is forced to learn more semantic information,which can help better suppress image artifacts in less angle scenes.In addition,we solve ring artifacts caused by detector element failure,and also compare the effects of different limited angle scanning ways on reconstruction images quality.Experiments show that the way of dividing all scanning angles into three sub-segments of average interval can obtain the best reconstruction results.