Efficiently train and validate a RapidPlan model through APQM scoring.

Area of Science:

• Radiation Oncology
• Medical Physics
• Computational Biology

Background:

• Knowledge-based planning (KBP) systems are crucial for optimizing radiation therapy plans.
• Current KBP training methods often involve time-consuming manual refinement.
• Objective patient-specific plan quality scoring can potentially streamline KBP model development.

Purpose of the Study:

• To propose and validate an intuitive method for training KBP systems.
• To utilize patient-specific plan quality scoring for KBP model training.
• To validate the performance of KBP models trained with the proposed method.

Main Methods:

• Eighty prostate cancer patient plans were ranked using the Adjusted Plan Quality Metric (APQM%).
• Two KBP models (APMQ25% and APMQ50%) were trained using libraries filtered by APQM% quartiles.
• Model performance was compared against Uncleaned and standard Cleaned RapidPlan models using closed- and open-loop tests.

Main Results:

• KBP models trained with APQM% thresholding demonstrated target coverage and organ-at-risk sparing equal to or superior than traditional methods.
• No significant performance differences were observed between the four tested models.
• APQM% analysis showed higher overall plan quality for APMQ50% and APMQ25% models compared to Cleaned and Uncleaned models.

Conclusions:

• Forward feeding KBP models with APQM% thresholding provides comparable or better results than manual refinement.
• A tighter APQM% threshold correlates with higher average plan quality.
• The proposed method offers a more intuitive and less time-consuming approach to KBP model refinement.