Automatic Treatment Planning For Cervical Cancer Radiation Therapy

In the workflow of clinical radiotherapy,the quality of radiotherapy plans determines the quality of patient treatment outcome.However,planning is a timeconsuming and labor-intensive process.Automatic planning can improve the efficiency of radiotherapy planning by learning the correspondence between patient anatomy and dose distribution.Current knowledge-based automatic planning methods mainly use low-dimensional features extracted from contoured structures to identify geometrically similar patients and then learn from existing clinical plans to predict a new plan.Here,we propose a knowledge-based planning method where the anatomical similarity is quantified by the rigid registration of the direct three-dimensional(3D)planning target volume(PTV)and organs at risks(OARs)between the incoming patient and database patients.We selected 81 cervical cancer IMRT plans from the first Affiliated Hospital of USTC and extracted the patient’s PTV,OARs and critical dose parameters.To test our strategy,a total of 20 patients were randomly selected and matched to the remaining database using leave-one-out method.The patient with most similar PTV and OARs of interest were chosen and the corresponding dose parameters were applied as optimization objectives.The generated automatic plans were compared with their clinical counterparts.The two major innovations are the direct application of the patient’s 3D anatomical structure registration for similarity match and the in-house customized scripts to automate the planning process.The results showed that automatic planning significantly reduced bladder and rectum’s V50Gy by 11.79±5.2%(p<0.01)and 2.85±3.16%(p<0.001)respectively while achieved similar PTV coverage and other OAR sparing.The entire optimization process,including both dose prediction and inverse optimization,takes less than 6 minutes.The above results show that automatic planning can efficiently generate radiotherapy plans of comparable or even better quality.Furthermore,we studied the feasibility of combining deep learning based automatic segmentation and the similarity match method to automate the entire planning process.Preliminary results show that the automated process can assist physicians and physicists to generate high quality plans.However,this is an exploratory attempt with limited data and needs further investigation on its clinical practicality.