Dose Distribution Prediction by Machine Learning (Random Forest): Impact of Data Specificity in Cervical Cancer IMRT

Abstract

Creating optimal radiotherapy plans is time-consuming and relies heavily on expert judgment to balance target coverage and Organ at Risk (OAR) protection. This study aims to fill the gap in existing radiotherapy approaches by integrating the Random Forest (RF) to predict dose determination in cervical cancer using Intensity Modulated Radiotherapy (IMRT). A retrospective analysis was conducted using 173 randomly selected cases and 102 specific data (stage I to IIIC1r cervical cancer, without prior surgery, and Whole Pelvic Non-Extended Field). Predictions were based on geometric relationships between organs and absorbed doses, with the model trained using decision trees and hyperparameter tuning via Random Search (RS). Model performance was evaluated using Mean Squared Error (MSE), Nash Sutcliffe Efficiency (NSE), and P-values. The ML-RF model showed improved performance with specific data, achieving the largest MSE reduction in bladder Dmax (0.017 Gy to 0.007 Gy). All parameters showed P-values above 0.05, indicating no significant differences in mean values between the predicted and clinical data. However, NSE values varied across parameters, with good performance for Right Femoral Dmax (NSE = 0.530) and Left Femoral Dmax (NSE = 0.554), while lower agreement was observed for PTV CI (NSE = –0.167) and PTV HI (NSE = –0.137), suggesting challenges in capturing distribution patterns. These results demonstrate the model’s ability to closely predict clinical dose distributions and underscore the value of incorporating specific clinical criteria. This approach may help reduce clinician workload and support planning standardization, while maintaining the importance of expert clinical judgment. Future studies should explore larger datasets, refined inclusion criteria, diverse treatment approaches, and beam weighting prediction to further improve model accuracy.