Research On CBCT/CT Alignment Algorithm Based On Super-resolution And Unsupervised Learning

At present,cancer is still one of the threats to human life,and adaptive radiation therapy is the most effective treatment option.In the treatment,the patient needs to take a planning computed tomography(p CT)scan to develop a treatment plan.Then,a set of cone-beam computed tomography(CBCT)scans were performed before each radiotherapy session and the specific radiation sites were delineated with alignment adjustments.However,the alignment algorithm currently used in clinical practice is very ineffective and requires doctors to manually perform alignment and radiotherapy area planning,which reduces the efficiency of adaptive radiotherapy.Therefore,a super-resolution-based unsupervised alignment method is designed to solve the above problem in this paper.To improve the accuracy of alignment in adaptive radiotherapy,two problems need to be solved.One is to improve the image quality of CBCT.The higher the image quality,the better the alignment effect.The second is to replace the existing rigid alignment network,whose single spatial deformation field is difficult to apply to diverse patient characteristics.First,this paper introduces a meta-learning method to propose a multi-scale super-resolution network that can improve the CBCT image quality to eliminate the artifacts and light spots in it.Second,this paper designs an unsupervised alignment network based on unsupervised learning,which only learns from the input image pairs and calculates the optimal spatial deformation field for each set of input image pairs.Finally,the unsupervised alignment network aligns the superresolution processed samples to improve the alignment accuracy of CBCT and p CT.To verify the improvement effect of each step,three rounds of experiments are conducted separately in this paper.First,clinical patient pelvic data are used to train and validate the super-resolution network to improve the CBCT samples,and then thoracic and liver data are used to verify the generalizability of the network.Then,the unsupervised alignment network is tested for the improvement of CBCT/CT alignment accuracy with the input of the original samples,and the generalizability of the network is verified on the thoracic and liver data.Finally,the improvement of the unsupervised alignment network is tested with the input of super-resolution-boosted CBCT samples,and the generalizability of the method is also verified on the thoracic and liver data.The experimental results show that the unsupervised super-resolution-based alignment method proposed in this paper can effectively improve the accuracy of alignment on different sites.