Feasibility of using geometry-based synthetic dose distributions for Knowledge-based planning model training

Knowledge-based planning (KBP) tools rely on large datasets of clinical plans, which are often difficult to collect, particularly in new or resource-limited centers. This study evaluates whether geometry-based synthetic dose distributions, generated from target and organ-at-risk (OAR) geometry, can serve as effective training data for RapidPlan (RP).

A software tool was developed to generate synthetic dose distributions based solely on target and OAR geometry. Synthetic doses were created for 25 head-and-neck (HN) cases (70 Gy/56 Gy/35 fractions) and 25 prostate cases (40 Gy/36.25 Gy/5 fractions). Synthetic RP models (synRP) were trained using these doses and compared to clinical RP models (clRP) from 171 HN and 97 prostate cases. Plan quality was evaluated using target coverage, OAR doses, and monitor units (MUs), while deliverability was assessed using portal dosimetry (3%/2mm).

Target coverage was equivalent between synRP and clRP. synRP reduced OAR doses: for prostate, bladder, and rectum doses decreased by 6% (p < 0.0001) and 15% (p < 0.00003), respectively; for HN, OAR sparing improved by 6.5% (p < 0.00005) for parallel OARs and by 4% (p < 0.005) for serial OARs. synRP plans required slightly more MUs (HN: 633 ± 55 vs 612 ± 62, p = 0.05; prostate: 2928 ± 181 vs 2883 ± 227, p = 0.3). Both plan types passed gamma evaluation (>98%).

Geometry-based synthetic dose distributions provide effective training data for KBP, enabling models that maintain target coverage while improving OAR sparing. This approach reduces reliance on large clinical datasets and offers a practical, resource-efficient strategy for establishing KBP models, particularly in centers with limited planning resources.