Study of BNCT Dose Prediction Based on Machine Learning

Boron dose accounts for over 70% of the total dose in boron neutron capture therapy (BNCT), serving as a critical factor in tumor eradication. Real-time and rapid acquisition of boron dose distribution in the target region during treatment is of paramount importance for accurately assessing therapeutic efficacy. This study employs machine learning techniques, integrating features of tumor location and boron concentration distribution, to achieve precise prediction of boron dose in the target area. Based on a radiation simulation human model with Chinese physiological characteristics, 20 brain glioma case samples were constructed. The Monte Carlo method was utilized to simulate the boron dose distribution under a posterior occipital irradiation scheme. The dataset was split into 80% training set and 20% validation set, and multiple machine learning algorithms (MLP, RF, LightGBM, CatBoost, XGBoost, SVM, GRU, and Stacking models) were applied to train and develop dose prediction models. Model fitting and predictive performance were evaluated using a series of metrics, including MSE, MAE, RMSE, RAE, MAPE, PCC, and R2. Results demonstrate that, in the task of predicting boron dose for complex tumor structures, the Stacking model outperforms others across all evaluation metrics, achieving an MAPE below 5% (meeting clinical requirements), an R2 of 0.9813, and an MSE as low as 0.0667, surpassing the predictive accuracy of other models. This study preliminarily validates the feasibility of machine learning in BNCT dose prediction, providing significant reference value for enhancing therapeutic outcomes and improving the precision of dose assessment in BNCT.