Research On The Scatter Correction Algorithm In Cone Beam Computed Tomography Imaging

X-ray Computed Tomography(CT)has become an indispensable solution in medical diagnosis and the imaging modalities changed with the development of system and scenarios.Cone Beam Computed Tomography(CBCT)that utilizes a large area flat panel detector was widely used in oral examination,3D digital angiography,Image-guided radiation therapy and other fields of application.However,the artifacts in CBCT images due to x-ray scatter interactions are more serious,which interferes the diagnostic.Various schemes have been proposed in literature in order to suppress the scatter artifacts.The approaches based on hardware led to additional cost and the effect of scattering cannot be sufficiently reduced.By analyzing the drawbacks of scatter estimation based on widely used pre-fitted scattering models,and the development of deep learning in the field of medical image processing,this thesis mainly focuses on exploring more accurate scatter correction methods.A two-stage kernel-based scatter correction method was proposed.Sparse and limitedangle re-projection simulation was operated by Monte Carlo on reconstruction images after the pre-fitted scattering model correction,in order to obtain more accurate and object-specific parameters of scattering models.The recalibrated models are then grouped and applied according to the projection approximation principle.The consistency deviation of scatter distribution estimated by pre-fitted scattering models on complex projections was reduced by the proposed approach.A quality assurance digital phantom was used to validate the performance.The error of restoration accuracy on CT number was reduced to 0.13%,the CNR of low-contrast resolution object was improved to 0.544,which is close to the reference image,and the uniformity difference was reduced to within 10 HU.Average estimation error of the anatomical regions with the most serious scattering in computational human phantoms is less than 15%.The effectiveness of the proposed method was also proved on clinical data.Based on the latest development of convolutional neural network on scatter estimation of x-ray images,a deep convolutional encoder-decoder based scatter estimation framework was designed in this thesis.A deep CNN was trained to yield more accurate scatter estimation by using simulated human phantom projections and scatter labels from Monte Carlo.This approach is not restricted to theoretical scattering models under specific scanning conditions,but adaptively learns the most suitable parameters from the observation data.The average estimation error is lower than 10% in all anatomical regions on simulation data,the difference of CT number between the corrected images and the reference is within 10 HU,and it is lower than 5HU on the head.During inference,the average computation time is less than 10ms/per projection,which means 5s for 601 projections.Qualitative and quantitative experiments were performed on simulated data and clinical data to prove the effectiveness.