Abstract: Objective In the preparation of CT-guided three-dimensional brachytherapy treatment plan for cervical cancer,the U-net model was used to complete the automatic segmentation of the applicator.Methods A deep learning model was created based on the U-net network.The data of 27 cervical cancer patients from December 2019 to October 2020 were preprocessed and written into the data set,which was divided into ratios of 15∶2∶10 for a training set,a verification set and a test set,respectively.The training set and verification set were put into the model for training and verification.The test set was applied to the trained neural network to segment the applicators.Dice similarity coefficient(DSC),95th percentile house dove distance(HD95),relative volume diffidence(RVD),precision and recall were used to evaluate this model.Results The average DSC of 10 patients in the test set was 0.90,HD95 was 1.26mm,RVD was-0.06,the accuracy rate was 0.94,the recall rate was 0.88,and the segmentation time was 5s.Conclusion In this study,the U-net network was used to realize the automatic segmentation of the applicator in the three-dimensional brachytherapy treatment plan for cervical cancer.It can be applied to the reconstruction of the applicator,which has great significance in realizing the automation of clinical planning.

Key words: Deep learning, Applicator segmentation, Brachytherapy, Cervical cancer