Explainable machine learning for patient-specific quality assurance in intensity-modulated radiotherapy based on...

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Patient-specific quality assurance (PSQA) is crucial for accurate intensity-modulated radiotherapy (IMRT) dose delivery.
• Conventional PSQA methods are labor-intensive and offer limited insights into gamma passing rate (GPR) variations.
• Anatomical features of planning target volumes (PTVs) and organs at risk (OARs) may predict GPR performance but remain underexplored in interpretable ML.

Purpose of the Study:

• To develop an interpretable machine learning (ML) framework for predicting IMRT GPRs.
• Utilize anatomical features from PTVs and OARs as predictors.
• Enhance understanding of factors influencing dose verification accuracy.

Main Methods:

• Retrospective analysis of 243 chest IMRT plans.
• Extraction of radiomic and dosimetric features from PTVs and OARs.
• Development of Random Forest and XGBoost models to predict GPRs under various gamma criteria (e.g., 3%/3 mm).
• Utilized Monte Carlo dose calculations for reference GPRs and SHAP for model interpretability.

Main Results:

• Both ML models demonstrated robust GPR prediction performance across structures and gamma criteria.
• Prediction errors decreased with less stringent gamma criteria.
• OARs showed higher prediction accuracy (e.g., heart MAE 0.06% at 3%/3 mm) compared to PTVs (e.g., MAE 1.98% at 3%/3 mm).
• SHAP analysis identified texture-based radiomic features as key predictors, with importance varying by organ and criterion.

Conclusions:

• Anatomical features can reliably predict IMRT GPRs using interpretable ML.
• The framework provides accurate predictions and mechanistic insights via SHAP.
• Offers a practical, transparent tool for improving IMRT PSQA and understanding dose verification variability.