Tilt Angle Estimation Of Cone Beam CT Detector Based On Deep Convolution Neural Network

Computerized tomography(CT)has become one of the most critical technologies in the field of biomedical and industrial nondestructive testing due to its advantages of non-invasive,high resolution,non overlapping and non-contact.Cone beam CT(CBCT)uses cone beam X-ray source and lat-panel detector,which can cover the measured object in a large range,collect the projection data in a short period of time,and reconstruct the clear three-dimensional image.So it is one of the main means of highresolution three-dimensional structural imaging.Filtered back projection algorithm is used in commercial CT system because of its fast calculation speed.CT imaging needs high-precision spacial positioning:firstly,the vertical line passing the center of the detector should also pass the focus of the x-ray point source and intersect with the rotation axis;Secondly,the rotation axis should be parallel to the detector plane.Any deviation of position or attitude will lead to reconstruction artifacts.The calculations of CBCT geometric parameters could be divided into analytic geometry algorithm and iterative geometry algorithm.The traditional analytic geometry algorithm needs a standard sample of precise geometry.Although the iterative geometry algorithm does not use any phantom,the calculation is cumbersome,even sometimes fall into local optimum.The main work of this paper is to estimate the tilt angle of CBCT detector,based on linear regression(LR),support vector regression(SVR)and the InceptionV3-R which based on the existing InceptionV3 algorithm,adding full connection layer.The experimental results show that the estimation result of InceptionV3-R algorithm is better than LR and SVR algorithm,and it can estimate the tilt angle of the detector with higher accuracy.The calculation period of angle estimation using the trained network is only 53 ms.The average absolute error of angle estimation based on InceptionV3-R is 0.0882°,the coefficient of determination R2-score is 0.9994.This paper for the first timedemonstrate the artifacts of tilt detector.Further more,the tilt angle estimation method based on depth convolution neural network is proposed.The feasibility of this method is verified by simulation.