Author: John Ginn, Zhuoyun Huang, Yongbok Kim, Ke Lu, Chunhao Wang, Yibo Xie, Zhenyu Yang, Jingtong Zhao 👨🔬
Affiliation: Duke University, Duke Kunshan University 🌍
Purpose: This study aims to develop and validate a machine learning model for predicting V60%, a critical dosimetric metric in LINAC-based Single-Isocenter-Multiple-Targets (SIMT) stereotactic radiosurgery (SRS). The goal of the model is to provide a reliable dosimetric outcome prediction for planning strategy decision-making and improve dosimetric quality consistency.
Methods: The developed model analyzes SIMT cases using 6 statistical features, including prescription level, target count, targets’ volumes, targets’ surface areas, targets’ equivalent spherical surface areas, and targets’ surface-to-surface distances. Gradient Boosted Trees (GBT) regression was employed for V60% prediction, with hyperparameters such as tree depth, learning rate, and feature splits optimized through grid search and 10-fold cross-validation. The training dataset consisted of 117 SIMT plans (2–53 lesions per plan, median = 6), with 27 independent cases used for testing (2-17 lesions per plan, median =4). All cases were generated by a single experienced planner. Model performance was first evaluated using metrics such as normalized mean squared error (MSE), and R².
Results: The achieved model demonstrated robust performance with a normalized MSE of 0.07, and an R² value of 0.99 within the test set. Predictions with uncertainty ranges were compared to ground truth values. The mean absolute difference (MAD) was 1.88cc, and the mean prediction uncertainty (MPU) of prediction was 3.35cc. Prediction times were less than one second per case. Residual and uncertainty analyses confirmed a strong agreement between predicted and ground truth values, demonstrating the model's reliability for clinical use.
Conclusion: The machine learning model for predicting V60% in SIMT radiosurgery was successfully developed and validated. By providing accurate and reliable dosimetric predictions, the model supports efficient treatment planning and reduces variability. Future work will focus on clinical integration and extending the approach to additional dosimetric metrics.