Cone-beam CT Guided Radiation Therapy Based On The Artificial Intelligence

Radiation therapy,surgery,and chemotherapy are three major ways of cancer treatment.More than 70% of cancer patients require radiation therapy.Radiotherapy aims to treat the tumors in the clinical target volume using radiation,and accurate beam delivery is critical to destroy tumors and protect surrounding normal tissues.However,due to positioning errors and the patients respiratory motion,tumor movement makes the target off the treatment beam,causing damage to normal tissues.Cone-beam CTguided systems use X-rays to monitor the tumor,making it possible to focus as much of the beam dose as possible on the tumor during treatment and minimize the damage to normal tissues.However,there are still three key issues that limit the accuracy improvement of cone-beam CT-guided radiotherapy: poor quality of cone-beam CT images,low accuracy of target volume localization,and difficulty of prognosis diagnosis.Based on these three key problems,we introduced artificial intelligence technology.We proposed: quantitative cone-beam CT imaging,precise cone-beam CTguided target volume localization,and intelligent prognosis diagnosis to improve the efficacy of cone-beam CT-guided radiotherapy.The purpose of the quantitative cone-beam CT imaging technique is to correct the scatter artifacts and ring artifacts of cone-beam CT.For the scatter artifact correction,we proposed a new "finger-crossing" blocker-based scatter correction method and quantitatively optimized the beam blockers geometric design with the help of a mathematical and physical model.It developed the scatter correction method based on the blocker from an empirical design to a theoretical design and successfully applied the blocker to clinical cone-beam CT scatter correction.We further reveal the blockers importance in clinical cone-beam CT scatter correction;we also propose a prior knowledge-based scatter correction technique for cone-beam CT.The results shown that the proposed method is better than the commercial cone-beam CT in the accuracy of the CT value.To improve the efficiency of scatter correction,we propose a conebeam CT scatter artifact correction method based on automatic image segmentation assistance,which does not need to change the systems hardware and scan protocol,and does not rely on prior information to achieve fast and accurate scatter artifact correction.By comparing with commercial cone-beam CT images,our method achieves better results.We propose an adaptive ring artifact correction method based on relative total variance to achieve the ring artifact correction without damaging the image spatial resolution for the ring artifact correction.The precise cone-beam CT-guided target volume localization technology aims to accurately and automatically contour the target volume and provide the key technical support for accurate radiation dose delivery.The development of deep learning has greatly advanced medical image processing.However,over-reliance on large data sets has become the biggest bottleneck for further development of deep learning.Unlike natural images,acquiring large amounts of medical images is more difficult.To alleviate the data dependence of deep learning models,we propose a one-shot learningbased cone-beam CT prostate localization model,which utilizes an elastic deformationbased data augmentation model and a dual residual network non-invasive accurate localization of prostate with only one patient data serving as the training set.Besides,we propose an unsupervised learning deformable registration framework based on narrow-band mapping for automatic contouring of the prostate target volumes.The proposed framework uses a regional deformable registration model to avoid image artifacts adverse effects and anatomical changes on target volumes.The intelligent diagnosis technique for radiotherapy prognosis is to automatically diagnose the status of patients after radiotherapy using deep learning.This paper constructs a multidimensional and hybrid long-short connection into a network for peritoneal metastasis prediction.The proposed Network introduces long connections between dense blocks to enable the network to extract peritoneal tumors multidimensional features.We joined several hospitals to evaluate CT images of 1978 patients.The model showed good agreement between the model prediction and the ground truth in external validation,substantially outperforming the traditional methods.Cone-beam CT-guided radiotherapy directly determines the effectiveness of radiotherapy,the patients risk of secondary cancer,and its prognosis of quality of life.This paper combines artificial intelligence to investigate quantitative cone-beam CT imaging,precise cone-beam CT-guided target localization,and intelligent diagnosis for radiotherapy prognosis.We validated our methods effectiveness through a large number of clinical cases,which is expected to provide key technical support for the realization of intelligent and rapid adaptive CT-guided radiotherapy.