KEY POINTS
- This retrospective, single-center study analyzed 243 thoracic intensity-modulated radiotherapy plans, including 100 lung cancer and 143 esophageal cancer cases, using structure-specific radiomic and dosimetric features.
- Random Forest and XGBoost models predicted gamma passing rates for the planning target volume, lungs, heart, and spinal cord under four criteria ranging from 3%/3 mm to 2%/2 mm. Reference values came from log-file-based Monte Carlo dose reconstruction rather than measurement-based quality assurance.
- Under the 3%/3 mm criterion, test-set mean absolute error was 1.98% ± 0.31% for the planning target volume, compared with 0.26% ± 0.04% for the total lung and 0.06% ± 0.01% for the heart.
- Prediction performance worsened under stricter criteria. For the planning target volume, mean absolute error increased to 4.97% ± 0.62% under 2%/2 mm, with systematic overestimation at lower gamma passing rates.
- Combining radiomic and dosimetric features produced the lowest overall error. Under 3%/3 mm, XGBoost root mean square error was 1.31% with combined features versus 1.37% with radiomics alone and 1.73% with dosiomics alone; SHAP analysis identified texture heterogeneity as a major predictor.
CLINICAL TAKEAWAY
Structure-specific machine learning may help identify plans requiring additional quality assurance and explain which anatomical or dose-distribution features influence predicted gamma performance. However, the model was trained at one institution, used calculated rather than measured dose verification, and was less reliable for targets and stricter gamma criteria, so external and measurement-based validation is required before clinical deployment.