Automated 4π radiotherapy treatment planning with evolving knowledge-base.

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Biology

Background:

• Non-coplanar 4π radiotherapy automates beam selection but requires complex tuning for optimal plan quality.
• Existing methods can be tedious and yield inconsistent results, necessitating more efficient automated solutions.

Purpose of the Study:

• To develop a fully automated 4π radiotherapy treatment planning system using evolving knowledge-base (EKB) planning guided by dose prediction.
• To improve organ-at-risk (OAR) sparing and overall plan quality in automated radiotherapy planning.

Main Methods:

• A statistical voxel dose learning model was trained on initial low-quality plans.
• A novel 4π optimization problem incorporated a one-sided penalty on OAR dose deviation from predicted doses.
• The fast iterative shrinkage-thresholding algorithm (FISTA) was employed for optimization over 10 EKB planning loops.

Main Results:

• EKB plans showed significantly higher plan quality metrics (PQM) for lung cases compared to manually created 4π plans.
• Head and Neck (HN) EKB plans achieved comparable quality to manual 4π plans but were surpassed by automated plans using high-quality training data.
• While individually evolved plans showed improvement, many were trapped in local minima, unlike the EKB approach.

Conclusions:

• Evolving knowledge-base planning offers a novel automated workflow for radiotherapy, leveraging predicted dose distributions to enhance plan quality.
• This technique can evolve from initially low-quality plans and effectively incorporate new beams for superior treatment outcomes.