Research On CBCT Synthetic CT Based On Cycle-consistent Adversarial Networks

In modern society,cancer has become one of the important diseases threatening human life,and adaptive radiotherapy is an important means of treating cancer.Patients are asked to take multiple Cone-Beam Computed Tomography（CBCT）and planning Computed Tomography（pCT）,which will be financially burdensome for the patient.The patient needs to be exposed to radiation when taking CT,which will cause greater secondary damage to the patient.Therefore,this paper investigates a synthetic CT（sCT）generated from the CBCT images method based on cycle-consistent Generating Adversarial Networks（cycle GAN）to address these problems.To address the problem that the CBCT images contains a large number of artifacts and cannot be used for image-guided treatment,this paper designs an improved experiment based on cycle GAN to synthetic CT from CBCT.First,the paper expands the CBCT/pCT dataset to increase the number of training sets.Secondly,the generator introduces improved skipping connections to improve the performance of the model.In comparison with the original cycle GAN and the cycle GAN with the attention gate,the quantitative and subjective experimental results show that the improved cycle GAN is superior.Finally,this paper tests the generalization ability of the head and neck model in the pelvic region,and the subjective and quantitative experimental results show that the robustness of the improved model is also better.In order to make full use of the image information of multiple CBCT scanned by the patient,this paper designs an experiment to synthesize CT based on multiple CBCT.The complementary anatomical structure information between multiple CBCT sets is used to enhance the quality of the sCT.Since some patients need to take 6 CBCTs in one treatment（e.g.,weekly CBCT）,4 models are designed in this paper.Model 1 is obtained by training CBCT and pCT in week 1,model 2 is obtained by training CBCT and pCT in two sets in weeks 1 and 2,model 3 is obtained by training CBCT and pCT in four sets in weeks 1 to 4,and model 4 is obtained from six sets of CBCT and pCT training from week 1 to week 6.The experimental results show that the quality of the sCT images by using multiple CBCT-trained models is better than that of the sCT synthesized by a single set of models（model 1）.