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

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Biology

Background:

• Knowledge-based planning (KBP) tools require large clinical datasets for training.
• Data collection is challenging, especially in resource-limited settings.
• This study explores synthetic data as an alternative for KBP model training.

Purpose of the Study:

• To evaluate geometry-based synthetic dose distributions as training data for RapidPlan (RP).
• To compare the performance of models trained with synthetic data (synRP) against those trained with clinical data (clRP).

Main Methods:

• Developed a tool to generate synthetic dose distributions from target and organ-at-risk (OAR) geometry.
• Trained synRP models using synthetic doses for head-and-neck (HN) and prostate cancer cases.
• Compared synRP models with clRP models using plan quality (target coverage, OAR doses, monitor units) and deliverability (portal dosimetry).

Main Results:

• Target coverage was equivalent between synRP and clRP models.
• synRP demonstrated improved OAR sparing in both prostate and HN cases.
• Both synRP and clRP plans achieved high deliverability, passing portal dosimetry evaluations.

Conclusions:

• Geometry-based synthetic dose distributions are effective training data for KBP.
• This method allows for KBP model development with improved OAR sparing and reduced dependence on large clinical datasets.
• Offers a practical and resource-efficient solution for KBP implementation, particularly in centers with limited data resources.