Can knowledge-based autoplanning keep up with advances in radiotherapy optimization for oropharyngeal cancer?

The performance of knowledge-based autoplanning depends on historical plans used to train the autoplanning models. Here, we investigate the applicability of an existing RapidPlan (RP) model for head and neck cancer to a novel planning and delivery solution (RapidArc Dynamic, RAD), which has the potential to improve plan quality through the integration of new degrees of freedom. RAD integrates static-angle modulated ports and a dynamic collimator in volumetric-modulated arc therapy (VMAT) fields.

A cohort of 48 oropharyngeal cancer (OPC) patients was retrospectively included in the planning study. Organ-at-risk (OAR) sparing was evaluated for VMAT and RAD, respectively, using both a VMAT-based RP model and a RAD-based RP model, resulting in four plans per patient: VMAT (RP-VMAT), VMAT (RP-RAD)), RAD (RP-VMAT) and RAD (RP-RAD). Differences were assessed with the related-samples Friedman's two-way ANOVA test by ranks, correcting the p value for multiple testing (p ≤ 0.05 considered significant).

For RAD plans, the RAD-based RP model improved the sparing of all OAR (p ≤ 0.001) except the parotids. However, the RAD plans were at least equivalent to the VMAT plans when optimizing with RP-VMAT, indicating the safety of initially implementing RAD with a VMAT-based RP model. In addition, when optimizing with RP-RAD, all OAR except the trachea were significantly better spared with RAD (RP-RAD) compared to VMAT (RP-RAD) (p = 0.039 for the esophagus and < 0.001 for the remaining OAR), with a median reduction of Dmean by 4.8 Gy and 3.5 Gy, respectively, for the larynx and the constrictor muscle. There was also a significant reduction in the estimated risk of dysphagia (1.9 pp) and acute mucositis (1.3 pp) (p ≤ 0.001).

VMAT-based RP models appear to remain applicable for optimization with the novel RAD solution until RAD-specific RP models are developed. Furthermore, RAD shows promise for OPC in terms of sparing of midline OARs.