Cone Beam X-ray Luminescence Computed Tomography Method Based On Residual Learning

As a new hybrid imaging technique,Cone beam X-ray Luminescence Computed Tomography(XLCT)can obtain both anatomical and functional information of the target,which has a broad development prospect in early diagnosis and targeted therapy of tumors.However,the inverse problem of cone beam XLCT is characterized by high ill-posedness and computational complexity,resulting in low spatial resolution and limited reconstruction accuracy.To solve this problem,a cone-beam XLCT neural network image reconstruction method based on residual learning was proposed in this thesis to improve the image quality of cone beam XLCT reconstructions.Firstly,a cone-beam XLCT quantitative phosphorescent distribution reconstruction method was constructed based on the residual network(Res Net)with rich characterization capability,which was used to simulate the nonlinear mapping between the photon density on the surface and the concentration distribution of the phosphorescent probe inside the imaging object.The simulation results show that the traditional method can not separate the adjacent targets with the center distance of 7mm,while the Res Net method can not only reconstruct the location,shape and concentration of the phosphorescent targets more accurately,but also improve the reconstruction speed greatly.However,it is still insufficient for the ResNet method to distinguish small targets.To solve this problem,this thesis introduced channel attention mechanism to allocate channel weight to the photon density features of cone beam XLCT,so as to enhance the characterization ability of Res Net.The simulation results show that the location error of the Res Net-CA method is as low as 0.23 mm for a small target with a radius of 1.5mm,which is about 0.2mm lower than that of the Res Net method.The similarity index between the reconstructed target and the real target is over 73%,which is 64% higher than that of the ResNet method.The physical phantom experiments further confirm the superiority of Res Net-CA method on the accuracy of target in cone beam XLCT.Finally,in order to improve the imaging efficiency and reduce the damage of biological radiation,Res Net method and Res Net-CA method were applied to study the image quality reconstructed with two projections.The results show that the acquisition time of two projections is 36% less than that of four projections.Meanwhile,the Res Net-CA reconstruction method can obtain a location error as low as 0.29 mm and a similarity index of more than 70%,while the image quality indexes of Res Net method are far inferior to that of the Res Net-CA method.Therefore,the Res Net-CA method can effectively improve the quality of cone beam XLCT image reconstructions.