Three-dimensional Dose Distribution Prediction Of Radiotherapy Based On Machine Learning

Radiation therapy is one of the main means of cancer treatment at present,and it’s key purpose is to reduce dose deposition of surrounding normal tissues as much as possible while ensuring that the target area reaches the prescribed dose.Dosimetric verification is also the main way of quality control and quality audit of current clinical radiotherapy technology.However,the quality of radiotherapy plans often depends on the experience of the plan designer,which will lead to a sub-optimal plan treatment for patients due to the uneven distribution of current medical level.At the same time,the clinical plan is subject to the unified standard which cannot provide individualized treatment plan for different patients.Studies have shown that modeling the effect of patient anatomical structure on dose deposition can predict the dose information of new patients before the plan design,which will provide standards for dosimetry verification and quality control,satisfy the individual specific needs of patients,and provide a basis for automatic radiotherapy.In this paper,we firstly study and analyzed the existing methods of dosimetry feature prediction,and comparing and summarizing the experimental ideas and implementation steps of different prediction targets such as Dosimetric Endpoints(DEs)prediction,DVH(dose-volume histogram)prediction,and three-dimensional dose distribution prediction.Among the above predicted targets,dosimetric endpoints is the prediction of one-dimensional dosimetric characteristics,and DVH is the embodiment of two-dimensional information,which contained information is not comprehensive enough to satisfy the clinical needs for 3d dose distribution information of some tumor types.Therefore,on the basis of the existing methods we firstly establish a 3D dose prediction model,which can prediction multi-organs simultaneously in a single model and capable of automatically learning the effect of patients’ geometric anatomical structure on dose distribution.This multitask relationship learning method was employed to model the relationships between multi-organs,based on the regularization framework.At the same time,a 3D dose distribution prediction model based on deep convolutional network is proposed to solve the problem about the loss of important information which cause by artificial feature extraction.The proposed model can learn the relationship between geometric structure distribution and patients’ three-dimensional dose distribution directly from the contour map of patients’ region of interest.Clinical radiotherapy plans(e.g.IMRT plans)treated on patients in the same tumor were collected and retrospectively studied.Our experiment used the spinal cord,brainstem and left and right parotid which were involved in the IMRT plan of nasopharyngeal cancer cases to test the multi-organs prediction model establishment.Total of 15 clinically treated IMRT plans were randomly collected in this study,and 10 of them were train data,and the other five were test data.Error between predicted percentage dose and clinical planned dose was calculated to verify the feasibility of the proposed method.The test results show a higher prediction accuracy and less data demand.And the average voxel dose error among spinal cord,brainstem,and left and right parotid was 2.01%±0.0249,2.65%%±0.0214,2.45%%±0.0217 and 2.55%±0.0216 respectively.At the same time,to test the feasibility of this model,a total of 37 cases of nasopharynx cancer were randomly selected from clinical planning,and 5 of them were used for model tests.The mean dose error and standard deviation of 5 test case is within 0.025 of the prescription dose,in where the average voxel dose error among Spinal cord,Brainstem,Parotid,Lens and PGTV was 2.04%±0.0154,2.56%±0.0226,2.07%±0.0188,1.16%±0.010 and 1.27%±0.10 respectively.Experimental representation the proposed mutil-task model can predict the dose of multiple organs more accurately in a single model.And proposed deep cornvolution network can predicted voxel dose distribution on head and neck cancer accurately.and analyze image feature automatically avoiding the artificial select geometric anatomical structure feature.