A practical method to quantify knowledge-based DVH prediction accuracy and uncertainty with reference cohorts.

Area of Science:

• Medical Physics
• Radiation Oncology
• Radiotherapy Planning

Background:

• Knowledge-based dose-volume histogram (DVH) prediction models are crucial for organ-at-risk (OAR) sparing in radiotherapy.
• Accurate quantification and interpretable error bands for DVH predictions are necessary for clinical adoption.
• Existing models' error bands may not always be directly interpretable as confidence intervals.

Purpose of the Study:

• To present and validate an independent error quantification methodology for DVH prediction models.
• To assess the predictive accuracy and error band interpretability of ORBIT-RT and RapidPlan models.
• To enable clinicians to independently evaluate DVH prediction models using local treatment plan data.

Main Methods:

• Developed an error quantification methodology using a local reference cohort of 45 prostate volumetric modulated arc therapy (VMAT) plans.
• Applied the methodology to DVH predictions from ORBIT-RT and RapidPlan models trained on 90 VMAT plans.
• Calculated prediction bias (μ), variation (σ), and root-mean-square error (RMSE) for each model and compared empirical RMSE to model-provided error estimates.

Main Results:

• ORBIT-RT demonstrated comparable or lower prediction bias and variation than RapidPlan for prostate OARs above 50% Rx dose.
• Above 80% Rx dose, both models showed low bias (<1%) and variation (<3-4%).
• ORBIT-RT's RMSE was slightly lower than RapidPlan's above 50% Rx dose, indicating improved accuracy.
• Model-provided error bands showed varying degrees of predictability, with ORBIT-RT's resembling an interquartile range and RapidPlan's being less predictive than the empirical RMSE.

Conclusions:

• The proposed methodology allows independent assessment of DVH prediction model accuracy using institutional data.
• Clinicians can use this method to interpret error bands and determine the potential for further OAR dose sparing.
• ORBIT-RT showed slightly superior predictive accuracy in the tested cohort, particularly at higher doses.