Research On CBCT Image Quality Improvement Algorithm Based On Deep Learning

Cancer is an important disease that causes a large number of deaths worldwide and poses a serious threat to people’s lives and health.Precision radiation therapy can effectively kill tumor cells and protect surrounding organs at risk,which can effectively control cancer progression.Cone-Beam Computed Tomography(CBCT)-based image-guided radiation therapy has the advantages of low radiation dose and low cost.However,CBCT image quality is low and there are more artifacts,which affects the accuracy of radiotherapy.Therefore,this paper conducts an in-depth study on improving CBCT image quality to address the problem of radiation therapy accuracy due to low CBCT image quality,and to improve the accuracy of adaptive radiotherapy.This paper first addresses the problems that affect the efficiency of experimental studies such as image backgrounds as well as size.Using image background removal,image alignment,image cropping with augmentation,and image block chunking with fusion methods.Increase the utilization of effective pixel points of images and reduce the experiment time to increase the research efficiency of CBCT quality improvement experiments.Secondly,to address the problem that the Cycle-consistent Generative Adversarial Network(Cycle-GAN)cannot sufficiently extract the high-level semantic features of images,which affects the accuracy of CBCT image quality improvement.Using the dilated convolution method,propose a Cycle-Residual Connection with a Dilated Convolution-consistent Generative Adversarial Network(Cycle-RCDC-GAN),and explore the results of CBCT quality improvement.In addition,the generalization performance of the model is explored: the pelvis region trained model is applied to the head and neck region;the head and neck region trained model is applied to the pelvis region;putting pelvic and head and neck images together to train the model and test.The experimental results show that the improved CBCT image quality using the 2D Cycle-RCDC-GAN method is clearer,the HU values are more accurate,and the model generalizability is stronger.Lastly,to address the problem that the 2D Cycle-RCDC-GAN improved CBCT image spatial structure discontinuity.The 3D Cycle-RCDC-GAN method is utilized and a 3D gradient loss function is added to enhance the structural continuity of the improved CBCT images.In addition,the generalization performance of the 3D model is explored,and the model trained in the pelvic region is applied to the head and neck region.The experimental results show that the improved CBCT images using the 3D Cycle-RCDC-GAN method have better structural continuity,more accurate HU values,and better model generalizability.