Automatic learning-based beam angle selection for thoracic IMRT.

Area of Science:

• Radiation Oncology
• Medical Physics
• Machine Learning

Background:

• External beam radiation therapy for thoracic cancer demands precise beam angle selection for optimal target coverage and minimal healthy tissue damage.
• Intensity modulated radiation therapy (IMRT) planning is complex and time-consuming, relying heavily on manual beam angle selection based on clinician experience.
• Automating beam angle selection can significantly streamline the IMRT planning process.

Purpose of the Study:

• To develop and evaluate a computationally efficient framework using machine learning for automatic selection of treatment beam angles in thoracic IMRT.
• To reduce the manual workload associated with IMRT planning for thoracic cancers.

Main Methods:

• A random forest regression algorithm was trained on a large dataset of clinically approved IMRT plans to correlate anatomical features with optimal beam angles.
• An optimization scheme was developed to select and adjust beam angles based on the learned relationships and interdependencies.
• The automated method was validated by comparing its selected beams against manually selected beams from clinical plans.

Main Results:

• The automated method demonstrated good correspondence with clinical beam angles (16.8° ± 10° angular distance, 0.75 ± 0.2 correlation).
• Generated IMRT plans achieved equivalent target coverage and organ-at-risk sparing compared to manually planned and clinical plans.
• 93% of automatically generated plans were deemed clinically acceptable by radiation therapy specialists.

Conclusions:

• Machine learning-based approaches are feasible for automating beam angle selection in thoracic IMRT planning.
• The proposed method shows potential to reduce manual planning efforts while preserving or improving treatment plan quality.