Purpose: Head and neck (HN) radiotherapy presents highly variable treatment geometries, which make optimal patient-specific organ at risk (OAR) dose objectives difficult to define. We sought to develop models of expected HN OAR doses based on geometric relationships between OARs and planning target volumes (PTV) to facilitate quantitative patient-specific plan quality evaluation.
Methods: Fifty-five treatment plans (VMAT or IMRT) from 41 patients were retrospectively analyzed. Primary and boost plans were assessed independently. For parotids, submandibular glands, and oral cavity, fractional volumes within 1, 5, 10, 20, 30, 40, and 50mm of the nearest PTV surface (including all dose levels) were computed for each plan. Linear, quadratic, and cubic polynomial fits were then generated to relate each distance metric with mean OAR dose (normalized to the lowest prescription dose in each plan). Optimal distance metrics were then selected for each OAR to minimize the mean absolute error (MAE) of the fit. The optimal polynomial order was set to 1 (linear) unless higher order models significantly improved MAE (P<0.05).
Results: The optimal parotid model (N=103 parotid contours including left and right) was a linear fit to the 10mm distance metric, which modelled mean parotid dose with MAE=4.2% (95% confidence interval, 95CI=0.2-10.6%). For oral cavity (N=55), the optimal model was a linear fit to the 20mm distance metric (MAE=6.9%, 95CI=1.1-15.1%) and for the submandibular glands (N=106), it was a quadratic fit to the 5mm distance metric (MAE=8.6%, 95CI=0.4-23.0%). MAE values did not change substantially (<0.4%) upon five-fold cross-validation. Target coverage was consistent across plans (D95% range: 98-103%, V95% range: 97-100%).
Conclusion: Simple models based on OAR-PTV distance metrics can produce accurate models of mean OAR doses in HN treatment plans. These models can be easily implemented in treatment planning systems to objectively assess plan quality against cases with similar treatment geometries.
Quality Assurance, Quality Control
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation