A Study On Automatic Planning For Intensity-modulated Radiotherapy Of Postoperative Esophageal Cancer Base On Deep Learning

Esophageal cancer has the characteristics of poor prognosis,high morbidity,high mortality and so on,which seriously threatens the health and life safety of Chinese residents.While intensity-modulate radiotherapy（IMRT）improves the conformity and uniformity of the target area,it can also effectively reduce the radiation dose to the normal tissues around the target area and indirectly increase the overall cure rate of the patient.It has become one of the main technical methods for the treatment of esophageal cancer.However,during the design process of IMRT plan,physicists need to adjust the objective function and weight repeatedly based on personal experience and continuously optimize to obtain the optimal plan,the design process is time-consuming,laborious,and subjectively,which greatly reduces the efficiency of plan design and the effect of radiotherapy.In order to solve the above problems,deep learning was adopted in this study.By building a deep learning model and relying on a large number of previous plan data in the tumor radiotherapy database of the hospital,we realized the3D dose distribution prediction for IMRT of postoperative esophageal cancer.At the same time,we tried to use this prediction model and Pinnacle~3Auto-Plan module to predict the 3D dose distribution of postoperative esophageal cancer,design the automatic plan respectively.Finally,the dosimetric comparison was made by comparing with manual plan,we could explore the dosimetric advantages of Auto-Plan for IMRT of postoperative esophageal cancer and whether the prediction model can guide the design of manual plan.Results:Study one:For the 10 testing cases,the average prediction bias ranged from-0.23%to 0.78%and the MAEs vary from 1.67%to 3.07%.The average DSCs value was above 0.91 for all isodose surfaces,especially,when the dose is less than 30Gy,the DSCs values above 0.95,the average HD₉₅varied from 0.51cm to 0.73cm.The dosimetric parameters of the prediction plan are all within the clinically allowable range and the relative dose deviation is less than 2%,there is no significant differences in other dosimetric parameters except for D₂of target area,D_maxof spinal cord and V₃₀of whole lung（P<0.05）.The predicted DVH curve and three-dimensional dose distribution are less different from the artificially designed plan.Based on the 3D U-Res-Net deep learning hybrid network,the dose distribution prediction of the intensity-modulated radiotherapy（IMRT）for the esophageal cancer after operation can be realized.Study two:The average DSCs value between the deep learning prediction plan and the manual plan is greater than 0.92 in isodose surface,and the average Hausdorff distance HD₉₅of the isodose surface is 0.58～0.62cm;The V₂₀,V₃₀,D_meanof total lung were slightly lower than those of manual plan（P<0.05）for the prediction model,meanwhile,the D₂,D₅₀,D_mean,HI of the target area and V₃₀of total lungs were better than those of manual plan（P<0.05）for Auto-Plan;Three-dimensional dose distribution of the three groups and the corresponding DVH curve showed that the three-dimensional dose distribution of the three groups had a little differences,and the DVH curves of the target area and organs at risk had a good agreement.The dosimetric results which predicted have certain advantages compare with the manual plan for lung tissue protection.When the Auto-Plan module is applied to the IMRT plan design of postoperative esophageal cancer,it can ensure that the organs at risk meet the clinical requirements and better dose uniformity.Conclusion:We could realize the dose distribution prediction of IMRT for the e sophageal cancer after operation,and the predicted dosimetric results have a cert ain advantages for manual plan and Auto-Plan in the lung tissue protection.