Prediction of Dose-Volume Histogram Parameters in Breast Cancer IMRT Using XGBoost from Geometric Features

Abstract

The planning process of Intensity-Modulated Radiation Therapy (IMRT) for breast cancer is time-consuming and susceptible to interplanner variability, compromising the reproducibility of treatment plans. Knowledge-based planning (KBP) offers a potential solution by predicting dose-volume histogram (DVH) parameters before optimization begins, yet most deep-learning implementations require volumetric CT data and substantial computational resources. This study developed an eXtreme Gradient Boosting (XGBoost) model using the top 50 anatomical geometric features, selected by importance ranking from a broader engineered feature set extracted from DICOM-RT Structure Set files, to predict 47 DVH parameters across six organs: planning target volume (PTV), heart, ipsilateral lung, contralateral lung, trachea, and spinal cord, without requiring CT voxel data. Data from 192 patients prescribed 50 Gy in 25 fractions were retrospectively collected. On an independent test set (n = 29), the model achieved a mean absolute error (MAE) of 1.53 Gy across all parameters, with 42 of 47 parameters (89.4%) below the 2 Gy clinical threshold. The PTV DVH Similarity Index reached 0.985 ± 0.007, the Homogeneity Index 0.189 ± 0.020, and the Conformity Index 0.910 ± 0.009, all within clinically acceptable ranges. Paired t-tests with Benjamini-Hochberg correction confirmed no significant difference between predicted and actual values across all 47 parameters (p > 0.05). Model accuracy was within the range of deep-learning-based KBP methods despite relying solely on tabular geometric features. These results support the XGBoost model as a computationally efficient KBP approach for breast IMRT planning in facilities without dedicated GPU infrastructure.