The First Affiliated Hospital of Soochow University Suzhou, Jiangsu Province
C. Ma1, J. Guo1, R. Cao2, Y. Wang3, L. C. Jia4, J. Zhou3, W. Zhang3, and J. Zhou1; 1the First Affiliated Hospital of Soochow University, Suzhou, China, 2Shanghai United Imaging Healthcare Co., Ltd, Shanghai, China, 3Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, 4Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China
Purpose/Objective(s): To design a trigger decision model based on automatic segmentation of organs at risk (OAR) for cervical cancer to assist doctors in making re-plan triggering judgments during the real-time online adaptive radiotherapy (OART), thereby shortening the OART practice time, reducing the waste of medical resources, and enhancing the consistency and accuracy of clinical practice decisions. Materials/
Methods: A total of 29 cervical cancer patients were enrolled in our department, 7 patients received radical radiotherapy and 22 patients received postoperative adjuvant radiotherapy, with a total of 646 treatment fractions. Divided all fractions into independent training sets and test sets at a ratio of 4:1. The United Imaging radiotherapy planning system was used to automatically outline the OAR in each treatment fraction, and the OAR and planning target volume (PTV) of the reference plan were copied to each treatment fraction through rigid registration. Referring to the triggering judgment ideas of senior radiotherapy doctors, 4 major categories of features were designed based on the volume or the single layer area of OARs, the centroid of OARs, the overlap change rate of OARs in treatment fractions and PTV in reference plan, and the Dice of OARs between treatment fractionation and reference plan. Feature extraction was performed on the three OARs of intestine, rectum, and bladder respectively, and a total of 49 features were obtained. Lasso was used to screen all features on the training set, a random forest classifier was constructed based on the filtered effective features, and the model performance was evaluated on the test set. Results: A total of 111 treatment fractions (17.18%) required replanning. In the training set, 10 effective feature categories were screened out, including the volume of bladder, the centroid of bladder, the Dice change rate of bladder, the volume of rectum, the Dice of rectum, the overlap change rate of rectum in treatment fractions and PTV in reference plan, the volume of intestine, the centroid of intestine, the area of intestine occupied by the target direction, and the overlap change rate of intestine in treatment fractions and PTV in reference plan. The model built based on the random forest classifier was subjected to 5-fold cross-validation, and the average AUC value was 0.96 and the average accuracy value was 0.90. Conclusion: This study was the first to construct a triggering decision model based on the features of OARs for cervical cancer during OART. This model could determine that whether re-planning should be triggered in OART. It had been proven to help physicians identify cervical patients who might benefit from replanning, thereby optimizing radiotherapy resource allocation and reducing radiotherapy toxic effects. In addition, since the construction of such features originated from the ideas of clinicians, these features were interpretable, which helped physicians to make clinical judgments and enhanced the accuracy and consistency of decisions.
