Using a commercial knowledge-based planning (KBP) DVH estimation tool, this study aims to investigate dosimetric outcomes between multiple in-house trained KBP models for lung VMAT radiotherapy plans considering laterality and outliers within each model. This study was conducted using planning data from Waikato Regional Cancer Centre (New Zealand) under approval from the Health New Zealand Te Whatu Ora Waikato research office and Waikato Māori research committee.

Six in-house KBP models were developed from 100 clinical VMAT plans with different PTV locations and prescriptions. An initial model was trained using all plans, becoming the ‘outliers- included’ model. A copy of this model was made in which vendor analytical tools were compared to an in-house developed Python script in order to judge and remove potential outliers, resulting in another ‘outliers-removed’ model. This process was repeated specifically for left-lung PTVs and the right-lung PTVs within the original training set, resulting in a total of six individual models.

These models were applied retrospectively to a different cohort of 15 VMAT plans where dosimetric outcomes were compared between each model and to the original clinical plan using the same field geometry.

All models performed similarly in OAR dose metrics. Interestingly, the left-lung trained models with and without outliers produced better PTV coverage regardless of the PTV laterality. Statistically significant differences (p < 0.05) between clinical and model-generated plans were observed for all PTV D₂% metrics.PTV homogeneity index was better in all outlier-removed models. Spinal cord Dmax was exceeded on one plan for both left-lung models (in a left-sided plan) despite having a lower mean than combined lung models.

The removal of outliers within the training set of a KBP DVH estimation model showed little effect on final OAR dose metrics, however helped increase PTV coverage and minimise cold/hotspots. Laterality-specific models did not yield conclusive differences, although the left lung models generally produced superior target coverage compared to the right lung models in all cases. A combined lung model with outliers removed and refined optimisation parameters, mainly for spinal cord Dmax, oesophagus Dmax and PTV D₂% is the most likely candidate for future clinical use in planning automation at Waikato Regional Cancer Centre.