Knowledge-based automated planning for oropharyngeal cancer

Area of Science:

• Radiation Oncology
• Medical Physics
• Computational Biology

Background:

• Automated treatment planning in radiation oncology aims to improve efficiency and consistency.
• Knowledge-based planning (KBP) leverages past treatment data to predict achievable dose distributions.
• Inverse optimization (IO) is a powerful technique for generating treatment plans based on desired objectives.

Purpose of the Study:

• To develop and evaluate an automated pipeline for generating radiation therapy plans for oropharynx patients.
• To combine knowledge-based planning (KBP) predictions with an inverse optimization (IO) pipeline.
• To compare two KBP methods, bagging query (BQ) and generalized principal component analysis (gPCA), for predicting dose-volume histograms (DVHs).

Main Methods:

• Developed BQ and gPCA methods to predict organ-at-risk (OAR) and target DVHs for oropharynx patients.
• Applied KBP models to a dataset of 217 patients using leave-one-out cross-validation.
• Integrated predicted DVHs into an IO pipeline to generate BQ and gPCA treatment plans; compared plans against clinical DVHs and clinical inverse optimized (CIO) plans.

Main Results:

• gPCA predictions closely matched clinical DVHs, while BQ predictions tended to be lower.
• BQ plans met clinical planning criteria for OARs more frequently than clinical plans (74.4%).
• gPCA plans satisfied target criteria most frequently (90.2%), outperforming clinical plans by 21.2%; gPCA plans maintained superior performance even when plan complexity was constrained.

Conclusions:

• An automated pipeline using DVH predictions can generate high-quality radiation therapy plans without human intervention.
• The gPCA-based automated plans demonstrated performance comparable to clinical plans, even when controlling for complexity.
• The BQ method showed underperformance compared to gPCA and clinical plans, particularly under complexity constraints.