Abstract
BACKGROUND: Online adaptive radiotherapy (OART) can address the reduction in radiation therapy precision caused by changes such as tumor shrinkage, rectal gas filling, and bladder distension. Currently, the determination of OART triggers during the fractions of radiotherapy requires clinicians to spend considerable effort and time on-site for assessments. OBJECTIVE: To reduce the clinical workload while maintaining the precision of radiotherapy, this study aims to explore the feasibility of developing a triggered OART prediction model for postoperative prostate cancer by combining radiomics and dosimetric features using machine learning. METHODS: A total of 244 fan-beam computed tomography (FBCT) images from 30 postoperative prostate cancer patients who received image-guided radiotherapy (IGRT) at least once per week at a single center were retrospectively collected. Two radiation oncologists (ROs) defined OART trigger labels based on morphological changes and dose differences between the current treatment fraction's FBCT and the planning CT for critical organs at risk (OARs). CT radiomics features and dosimetric features were calculated for five regions of interest (ROIs): clinical target volume (CTV), planning target volume (PTV), PTV minus CTV (PTV-CTV), bladder, and rectum. The least absolute shrinkage and selection operator (LASSO) algorithm was used for feature selection. Support vector machine (SVM) algorithms were employed to establish CT radiomics feature models (R models) and dosimetric feature models (D models), resulting in 10 OART trigger prediction models. Finally, we utilized model integration methods to create a comprehensive predictive model that integrated the 10 OART trigger prediction models, achieving optimal performance. Five-fold cross-validation was used for preliminary evaluation of the model performance, and further assessment was conducted on an independent test set. RESULTS: Eight hundred fifty-one CT radiomics features were extracted from each ROI. Additionally, 14 dosimetric features were calculated for the target-related ROIs (CTV, PTV, and PTV-CTV), and 9 dosimetric features were calculated for the OARs (bladder and rectum). Among the CT radiomics models, the PTV-CTV-R model had the highest AUC at 0.824; among the dosimetric models, the Bladder-D model had the highest AUC at 0.85. The comprehensive predictive model for triggered OART (Comprehensive-E model), which integrated all radiomics and dosimetric features, achieved an AUC of 0.893, significantly outperforming models that used only CT radiomics or dosimetric features. CONCLUSION: This study is the first to combine radiomics and dosimetric features using machine learning to establish a triggered OART prediction model, demonstrating the feasibility of using CT radiomics and dosimetric features for OART trigger prediction. By employing model integration techniques, we addressed the issue of one-dimensional feature bias, significantly enhancing model performance. This research provides new data analysis methods for adaptive radiotherapy triggering in postoperative prostate cancer, offering substantial clinical value.