Author: Lian Duan, Stephen F. Kry, Hunter S. Mehrens, Paige A. Taylor 👨🔬
Affiliation: The University of Texas MD Anderson Cancer Center, UT MD Anderson Cancer Center 🌍
Purpose: To develop a machine learning model for predicting dose delivery accuracy and identifying its key factors in IROC’s proton phantom program.
Methods: IROC’s proton QA program has six proton phantoms: Brain(N=59), Head and Neck(N=65), Lung(N=91), Liver(N=76), Prostate(N=75), and Spine(N=37). Phantom irradiations from 2009-2024 were analyzed using three machine learning algorithms: k-nearest neighbors (kNN), support vector machines (SVM), and random forest. Irradiation results included pass/fail classification, average TLD ratio for targets and OARs (if applicable), and percent of pixels passing film gamma analysis. The following input categories of treatment parameters were used: CT-stopping power calibration (8 phantom materials & 4 linear spline fit), machine output, patient-specific QA, dosiomic features, and DVH metrics. Input features for the machine learning models were reduced by a voting feature selection.
Results: For predicting pass/fail in all phantoms, random forest sensitivity (75±8%), or the identification of failure phantoms, and kNN (81±10%) were comparable, both outperforming SVM (69±8%). Accuracy showed similar trends. Averaged across all phantoms, the root mean square error between prediction and observation showed that random forest (TLD: 0.03±0.02,gamma: 9.7±4.1%) was comparable to kNN (TLD: 0.04±0.02,gamma: 12.1±5.8%) and SVM (TLD: 0.04±0.02,gamma: 12.2±6.7%) for target TLD and the percentage of pixels passing gamma. For random forest, the most important factors were target and OAR dosiomics features and OAR DVH metrics. CT-stopping power table, machine output, and patient-specific QA had minor influences on prediction.
Conclusion: This study demonstrates that predicting proton phantom performance with machine learning is feasible across all phantoms. Random forest and kNN were the most successful machine learning techniques, and dosiomics was the most important feature for prediction. This approach shows promise for incorporating machine learning techniques into QA processes and optimizing treatment planning in proton therapy. This is progress along IROC’s goal of supplementing its existing physical phantom program with computational techniques.