A phantom geometry approach to develop a RapidPlan knowledge-based planning model.

Knowledge-based radiation therapy (KBRT) models, including RapidPlan (RP), significantly improve treatment planning by enhancing both consistency and plan quality. This study focuses on the development and validation of an RP model utilizing phantom geometries that simulate prostate patient anatomies for optimized dosimetry. A library of 240 volumetric modulated arc therapy (VMAT) plans was developed using customized phantom geometries to simulate prostate anatomy with varying sizes for the planning target volume (PTV), bladder, and rectum. All plans were optimized following standard clinical constraints that ensure at least 95 % target coverage while adhering to the ALARA principle for organ protection. RP model was validated on data from 30 prostate cancer patients with three dose regimens: 50 Gy in 25 fractions (reference), 60 Gy in 20 fractions, and 78 Gy in 28 fractions. Plans for the 60 Gy and 78 Gy cohorts were compared with and without additional optimization objectives to evaluate improvements in sparing organs at risk (OAR). The RP model, trained on 240 phantom plans, showed strong target coverage, with 95 % of the PTV receiving at least 95 % of the prescribed dose (mean PTV dose: 48.25 ± 0.13 Gy in phantoms; 48.02 ± 0.26 Gy in patients). For the 78 Gy in 28 fraction prescription, rectal sparing was significant, with Rectum V65Gy (%) decreasing from 15.4 % to 10.1 % (p = 0.001) and V40Gy (%) from 32.6 % to 24.5 % (p = 0.004). In contrast, bladder dose reductions were minimal and not statistically significant. For the 60 Gy in 20 fraction prescription, bladder sparing was notable, with Bladder V48Gy (%) decreasing from 16.8 % to 9.7 % (p = 0.005) and V56.8 Gy (%) from 11.4 % to 5.0 % (p = 0.009), while rectal dose improvements remained modest and non-significant. Significant changes were defined as p < 0.05. The RP model, trained on a library of phantom-based plans aligned with body dimensions and dose constraints, demonstrated strong target coverage and effective sparing of organs-at-risk (OAR) across various prescription regimens. Validation with 30 prostate cancer patients confirmed that this method provides clinically relevant dose predictions. Though phantom datasets cannot fully replicate the complexities of patient anatomy, this approach enables scalable training of knowledge-based planning (KBP) models, reducing variability among planners and maintaining plan quality when real patient data is scarce. This RP model, based on phantom geometries, delivers accurate dosimetry predictions while meeting clinical objectives, indicating significant potential for enhancing prostate cancer treatment planning.