ESTRO 2025 - Abstract Book

S3677

Physics - Quality assurance and auditing

ESTRO 2025

Conclusion: We have demonstrated that it is possible to perform patient QA by faithfully tracking the actual trajectory with high accuracy using conventionally used radiochromic film, ionization chamber, and 4D-moving phantom. In the future, we plan to compare the results of this film and chamber study with those of Delta4 HexaMotion (ScandiDos).

Keywords: Synchrony, Respiration tracking, 4D moving phantom

References: 1 ） Ferrisx WS, Culberson WS, Smilowitz JB et al., Effects of variable-width jaw motion on beam characteristics for Radixact Synchrony, J. Appl. Clin. Medical Phys., 22(5), 175-181, 2020 2 ） Ferris WS, Kissick MW, Bayouth JE et al., Evaluation of radixact motion synchrony for 3D respiratory motion: Modeling accuracy and dosimetric, fidelity J. Appl. Clin. Medical Phys., 21(9), 96-106, 2020 3) Chen Q, Rong Y, Burmeister W et al., AAPM Task Group Report 306: Quality control and assurance for tomotherapy: An update to Task Group Report 148, Med. Phys., 50(3), e25-e52, 2023

3817

Digital Poster Predicting IROC’s Thoracic Dosimetry Audit Performance Using Random Forest Modeling and Feature Interpretability. Hunter Mehrens, Stephen F Kry Radiation Physics, MD Anderson Cancer Center, Houston, USA

Purpose/Objective: To identify the factors that predict performance on IROC’s thoracic dosimetry audits.

Material/Methods: To pass IROC’s heterogeneous moving lung phantom audit, TLD in the island lung tumor must agree with the TPS calculation within -8% and +5%, and ≥ 80% of pixels must pass a 7%/5mm gamma analysis on film. 1,521 thoracic phantom irradiations from 2014-2022 were reviewed. Random forest modeling was used to predict phantom performance based on: (1) complexity metrics, (2) machine output, (3) Institutional IMRT QA, (4) size of delivered dose cloud (in direction of motion), (5) TPS parameters, (6) dosiomic features, and (7) DVH metrics. Feature selection was performed using a voting scheme to reduce the number of input features. Additionally, Shapley Additive Explanations (SHAP), a machine learning interpretability algorithm, was employed to assess the influence of features on predictions. Modeling aimed to predict phantom performance across 14 stratified treatment delivery combinations (different TPS, linac types, beam energies, and motion management combinations). Results: Averaged across the 14 models, AUC for predicting pass/fail was 0.96 ± 0.04 and mean absolute error between the predicted and measured was 0.012 ± 0.003 for TLD and 3.3% ± 1.5% for gamma analysis. For pass/fail prediction, complexity metrics, dosiomics features, and DVH metrics were the most influential. For gamma analysis, institutional IMRT QA also played a significant role. For TLD, all parameter sets had approximately equal influence. However, a deeper analysis of specific models revealed that different models were influenced by different input parameters. For example, in the TrueBeam-Eclipse-AAA-VMAT-6 MV model, ITV motion management was most influenced by DVH metrics, while gating motion management was influenced by complexity metrics. Furthermore, even when similar classes of features (e.g., complexity metrics) were important, the specific complexity metric that predicted phantom performance varied between models, reflecting differences in how planning systems manifest plan complexity.