Research On Scatter Correction Method For Cone Beam CT Based On CNN And Transformer

Cone-beam computed tomography(CBCT)is playing an increasingly important role in image-guided radiotherapy.However,scatter of X-rays through objects during CBCT imaging can cause CBCT images with artifacts and noise,which reduces the image quality and diagnostic usability of CBCT images and hinders the further application of CBCT in radiotherapy.Therefore,it is the ultimate goal of CBCT scatter correction to remove the effect of scatter on CBCT images and restore the low quality CBCT images contaminated by scatter to high quality images.In this thesis,we propose two CBCT scatter correction methods based on existing deep learning techniques and validate their considerable correction ability on abdominal CBCT datasets,respectively.The main work of this thesis is summarized as follows.(1)Most current advanced scatter correction methods using deep learning are based on convolutional neural networks(CNNs),and few have attempted to use the Transformer,which has shown impressive performance in advanced vision tasks.This thesis proposes a U-shaped CBCT image scatter correction model,FSTUNet,which uses CNNs to extract shallow texture details of CBCT images and enhances the grasp of deep global features using this thesis’ new modified structure of Swin Transformer,Flip Swin Transformer.Through extensive experiments on the Monte Carlo simulated dataset and frequency split dataset,as well as comparison tests with other current deep learning scatter correction methods,the root-mean-square error of images corrected with this method was reduced from more than 100 H to about 7H,and both the structural similarity index measure(SSIM)and the universal quality index(UQI)were close to 1.This thesis demonstrates its efficient scatter correction capability and considerable clinical feasibility.On the other hand,to demonstrate the important role of global features for scatter correction,this paper also uses the division of images into image blocks with different resolutions to verify the law that the RMSE decreases from 69.95 HU to 28.24 HU,i.e.,the correction ability gradually increases,as the resolution increases,i.e.,the global features are enhanced.Finally,the image preprocessing also affects the final correction results,and this thesis make a detailed comparison between the two data processing methods of maximum-minimum normalization and no normalization in the current study,and show the intuitive effect of normalization and no normalization on the correction results at the pixel level.(2)Traditional block-based methods can obtain scatter truth values but destroy the anatomical structures in the projection.Given the amazing capability of current deep learning techniques in the field of image restoration,it is expected to present more considerable scatter correction capability if deep learning techniques are combined with blocking techniques.In this thesis,a new method for block-based scatter correction in the projection domain using UNet optimization is proposed.On the one hand,this thesis makes full use of the true value of scatter under the block,and then realize the scatter prediction beyond a single technique by the correction of UNet.On the other hand,this thesis takes advantage of the highly redundant information of adjacent angle projection and adopts the complementary technique of adjacent angle projection,and then through UNet’s excellent image restoration capability to achieve considerable projection restoration.Finally,in both projection and image domains,the results obtained using this method are compared with two techniques based on pure deep learning,i.e.,direct prediction of scatter-free projections and indirect prediction of scatter signals using neural networks,and pure blocking-based techniques.The method reduced the RMSE to 15.52 HU,which is still 12 HU lower than the most effective current method,demonstrating the advantages of this combined method.