Machine learning approach to estimation of in vivo measured dose and treatment planning for total body irradiation

Abstract

Purpose This study aims to enhance the conventional Total Body Irradiation (TBI) dose calculation method which relies on manual calculation using tissue maximum ratio, off axis ratio, and scatter factors, by incorporating machine learning (ML) to predict in vivo measured doses and optimize the number of compensators needed, thereby improving the accuracy and efficacy of TBI treatments. Materials and methods A retrospective analysis was performed on patient data between 2018 and 2022, involving 96 patients. The data included demographics, treatment specifics, and in vivo dose measurements. We developed and evaluated various ML models, including Random Forest, XGBoost, LightGBM, and Gradient Boosting Regressor, using 70% of the data for training and 30% for testing. Model performance was assessed using the Mean Absolute Percentage Error (MAPE). The optimization process adjusted compensator numbers based on ML predictions to better align with the prescribed dose. Results All ML models outperformed the conventional method, with Gradient Boosting Regressor demonstrating the best performance. It achieved an overall MAPE of 2.52% (maximum 2.95% at hip with hand, minimum 2.19% at umbilicus without arm) in predicting measured doses, which is a significant improvement compared to the conventional method's overall MAPE of 4.01% (maximum 8.69% at head, minimum 2.74% at umbilicus without arm). Furthermore, the optimization process successfully generated treatment plans with a high degree of conformity to the prescribed doses, achieving a MAPE of 1.69% between the optimized and prescribed doses. The head region, which exhibited the largest discrepancy in the conventional method, saw substantial improvement, while the neck region required an increase in compensators to correct underestimation. Conclusion The machine learning approach successfully improved the accuracy of in vivo dose predictions and compensator optimization, suggesting an enhancement of conventional methods. This advancement holds the potential to deliver more precise and personalized radiation therapy in TBI treatments.