The Research On Respiratory Motion Prediction Algorithm Based On Real-time Tracking Technology

Radiotherapy remains one of the three major treatments for cancer patients,with domestic data showing that 70%of cancer patients have received radiotherapy.The radiotherapy process contained numerous links that might cause errors,is challenge of radiotherapy currently,especially in the radiotherapy of thoracic and abdominal tumor.The respiratory motion affected the thoracic and abdominal organs,the movement may lead to tumor target irradiation and the overflow of normal tissue around the tumor target caused the target under the dosage but increase the surrounding normal tissue or organ dose,which leads to the side effects of radiotherapy.So a new technique is being developed that can accurately compensate for the movement of the target area caused by respiratory movement,while ensuring that the tumor is receiving enough,and the normal tissue around it is as small as possible.The effect of respiratory movement on radiotherapy is mainly reflected in three aspects:(1)the organ movement caused by respiratory movement and(2)there is an error between the image structure and the anatomical structure,and(3)the dose distribution is deformed.Real-time tracking technology can not only let the patient breathe freely,but also through the noninvasive identification to find the relationship between signal and signal in vivo in vitro,achieved by changing the irradiation or reverse adjust treatment bed relative position to shoot wild the center and the target area remains unchanged.At present,real-time tracking technology is the most promising research direction to solve the respiratory motion problem in chest and abdomen tumor radiotherapy.Real-time tracking is based on the relationship between in vivo and in vitro signals.Firstly,the in-vivo and in vitro data were collected synchronously for a period,and the relationship between them was obtained by training,that is,the correlation model.Then combining the model with the obtained vitro data to calculate the position of the tumor in the body,and adjusting the irradiation field or the treatment bed to compensate the error caused by breathing motion.There are two important algorithms:respiratory motion prediction algorithm and association algorithm.Respiratory motion prediction is due to hardware response and other reasons;the entire system will inevitably produce a certain delay,so we must first predict the external signal to compensate for the system delay.This paper presents the prediction algorithm and correlation model of respiratory motion based on support vector regression.The basic ideas based on support vector regression algorithm is selection of history data as the training set at the first,then selected the appropriate mapping relationship between input and output,kernel function,optimization method that is suitable to obtain the optimal model parameters,obtained the regression model by training.When there is a new signal to respiratory movement,through the regression model to calculate the predicted value.The key problem is determining the size of training set,the number of feature vectors and another problem is how to find the optimal parameters.In this paper,the penalty function C and the insensitive loss function ε are determined based on the training set data and the noise level proposed in the literature.In the process of SVM,the new data is added to the training set,that is to say,every time there is new data,it needs to be retrained.To solve the problem,accurate online support vector regression(AOSVR)was applied,this method is based on the traditional method to realize the online learning and updating the training model,the difference with traditional models is that when data is updated,SVR model is not needed to be retrained.Instead,a data is added directly or a data is removed,and the characteristics of SVR model and corresponding training set data are adjusted online,so that the model is still in line with the prediction conditions.In order to demonstrate the performance of the above method,the vivo and vitro synchronization data from the Institute of Cognitive and Robot Studies at the University of Rubek,Germany,were used for the experiment.To verify the performance of respiratory motion prediction algorithm based on support vector regression(SVM),the paper first predicted in vitro data of seven samples.At the same time,it used K-nearest neighbor regression and linear regression and the traditional support vector regression(GA-SVM)is compared.The results showed that when the delay is 300 milliseconds,the mean absolute error(MAE)of K neighbor regression,linear regression,GA-SVM and AOSVR are mean 1.24 mm,0.44 mm,0.42mm,and 0.30mm.And the mean value of Root Mean Squared Error(RMSE)is 1.67 mm;0.60 mm,0.88 mm,0.37 mm respectively.From the evaluation index,compared with linear regression,GA-SVM is superior to linear regression,which fully reflects the strong robustness of support vector machine.AOAVR is compared with GA-SVM,the same sample has good performance in the time-consuming aspect of the algorithm,so online support vector regression solves the problem of long time consuming in the traditional off-line model.Support vector machine is used to the association model by using in vivo and in vitro signal to find the relationship between the signals.The average nRMSE of the seven sample correlation models is 0.51,which can meet the needs of practical applications.But support vector regression(SVM)for association model needs further optimization to make it more practical.At the end of the paper,we make a summary for our study work,to find the shortcoming in the research,and introduce the future work.