Knowledge-based treatment planning in breast cancer radiotherapy: comparing different machine learning algorithms.

Abstract

Knowledge-based planning (KBP) has emerged as a promising approach to improve the consistency and efficiency of radiotherapy by utilizing prior clinical data to guide treatment planning. This study aimed to evaluate the performance of various machine learning (ML) algorithms for predicting key dosimetric parameters in 3D conformal radiotherapy (3D-CRT) of left-sided breast cancer, with the goal of reducing manual trial-and-error during plan optimization. A retrospective dataset of 75 breast cancer patients treated with 3D-CRT was used. Geometric and dosimetric features, including beam and organ-at-risk (OAR) parameters, were extracted from clinically approved plans. Ten supervised regression algorithms, including SVR, LASSO, Ridge, AdaBoost, Gradient Boosting Regressor, Histogram-Based Gradient Boosting, KNN, MLP, SGD, and Kernel Ridge Regression, were trained to predict dosimetric outcomes such as D_mean and V_x for the heart and lung, and homogeneity index (HI). Data preprocessing included outlier capping, quantile-based oversampling, and MaxAbs scaling. Model performance was evaluated using fivefold cross-validation and an independent test set, employing RMSE, MAPE, and MedAE as metrics. Among the evaluated algorithms, the KNN model demonstrated the consistent predictive performance across all dosimetric endpoints, achieving the lowest RMSE and MAPE for D_mean Heart (2.09 Gy, 0.17) and D_mean Lung (3.12 Gy, 0.15). Feature importance analysis identified geometric parameters such as borders, wedge angles, and OAR volumes as the most influential predictors. ML-based KBP can accurately predict dosimetric outcomes prior to dose calculation, improving planning efficiency and consistency in breast 3D-CRT. The KNN algorithm showed the highest reliability, suggesting its suitability for integration into clinical decision-support systems.