KEY POINTS
- This retrospective planning study developed a locally trained RapidPlan knowledge-based planning model for head and neck, brain, and central nervous system radiotherapy using 497 training plans from an initial 594-plan dataset.
- Model performance was evaluated in 370 paired plan comparisons using identical beam geometry; 303 validation plans overlapped with the training model, while 67 were plan-level held-out cases.
- RapidPlan maintained broadly comparable target coverage, with no significant difference for PTV1 D98% and only small statistically significant decreases for PTV3 and PTV4 D98%.
- Organ-at-risk sparing improved for several structures, including oral cavity Dmean −7.62%Rx, larynx Dmean −7.57%Rx, salivary gland Dmean −4.36%Rx, and brain Dmax −5.80%Rx.
- Plan complexity did not increase: monitor units decreased from 764.2 ± 275.5 to 695.8 ± 210.3, delta −68.4 MU, p < 0.001.
CLINICAL TAKEAWAY
This study supports knowledge-based planning as a standardization and decision-support tool for complex head and neck and brain radiotherapy workflows. The useful message is that moderate model statistics, such as R², did not prevent clinically meaningful organ-at-risk sparing when the training library was carefully curated. The evidence remains technical and institution-specific, limited by retrospective design, substantial training–validation overlap, and lack of fully patient-independent external validation.