Three-dimensional Dose Distribution Prediction Of Radiotherapy

Radiation therapy is one of the three major means of cancer treatment,its basic objective is to maximize the treatment gain ratio,which is to ensure that the target area reaches a certain coverage when the prescription dose is irradiated,while the normal tissues are protected from or unnecessarily irradiated.However,being objectively limited,the actual radiotherapy plan performed by the clinical patients is often difficult to achieve the optimal homogenization.The reason is that the current plan design process using trial-and-error method to continuously adjust the plan optimization goals/constraints,and the plan quality depends on the degree of experience of the plan designer and the time and energy spent on it.Furthermore,clinical planning is designed to meet the relevant criterions.Because of the existence of individual differences,meeting the basic criterions does not mean that the individual case is optimal.Therefore,accurately predicting patients' corresponding quality plan dosimetry goals and fully reflecting the specificity of patients before treatment plan design can not only provide a precise measurement for the quality control of the individualized planning design,but also provide a good start for the optimization of intelligent plan.Aiming at the prediction of the dosimetric characteristics of radiotherapy planning,this paper introduces the core ideas and implementation steps of different methods for different prediction methods of dosimetry characteristics of radiotherapy plans.The predicted dosimetry features include the Dosimetric Endpoints(DEs),the DVH(dose-volume histogram)and the three-dimensional(3D)dose distribution.The DEs is predicted by establishing a linear correlation model between dDTH(OVH or DTH differential forms)and DEs values.There are three ways to predict DVH.A model is estimated based on the sub-DVH probability,a model is established based on the estimated dose influence probability distribution and dose probability distribution,and a model is established based on support vector regression(SVR).The three-dimensional dose distribution prediction method is using support vector regression or artificial neural network(ANN).The two method of prediction is to construct the correlation model between the information and the dose distribution extracted from the radiotherapy plan.In this paper,the advantages and disadvantages of various dosimetric characteristics prediction methods are analyzed in detail.Based on the comparative analysis,the 3D dose distribution prediction model of radiotherapy plan was established by using SVR and ANN learning methods respectively.To verify the dose prediction method,a number of cases of prostate cancer IMRT/SBRT plan were collected,and the corresponding spatial location information,volume information and voxel dose were extracted from clinical data,so as to establish the dose prediction model.The SVR learning method and kernel function are applied to construct the dose prediction model.The dose difference of training set percentage value is no more than 1%,and the dose difference of prediction set percentage value is not more than 5%.The prediction performance of the model is good.A three-dimensional dose prediction model was constructed by feed-forward back propagation neural network.The result of model training shows that the difference between the dose difference of training set,that is,the difference between the predicted dose and the dose of clinical plan dose is less than 1.7%,and the dose difference of training set is small.Compared with the clinical plan DVH,the model prediction DVH has little difference,the difference of DVH is within 2%,and the data fitting effect is ideal.In validation samples,the dose difference was within 5%,and the difference of DVH was less than 3%.Three dimensional dose distribution shows a small dose difference,and the prediction dose distribution is reasonable.The error distribution is concentrated near the 0 value,showing symmetrical distribution,and the prediction result is good.The prediction result of the bladder model was the best,and the rectum and the femoral head model were the next.This paper used support vector regression and neural network learning method to establish the dose prediction model between the same type of tumor patients with geometric characteristics of anatomical structure and the corresponding three-dimensional dose distribution plan.The establishment of the correlation model can not only provide the quality standards for the plan,but also serve as a good initial value for subsequent optimization,and lay a foundation for realizing automatic plan control and plan design.