Automatic Treatment Planning for Radiation Therapy: A Cross-Modality and Protocol Study

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Healthcare

Background:

• Automated treatment planning aims to improve efficiency and consistency in radiation therapy.
• Cross-modality transfer learning is an emerging area for adapting AI models to different treatment techniques.
• Multicriteria optimization (MCO) allows for balancing competing clinical objectives in treatment planning.

Purpose of the Study:

• To assess the cross-modality applicability of 3D dose predictions from a model trained on intensity modulated radiation therapy (IMRT) data.
• To evaluate the impact of integrating a multicriteria optimizer (MCO) for adapting dose predictions to institutional preferences.
• To explore the potential for reducing plan generation time and variability in radiation therapy.

Main Methods:

• A 3-stage U-Net model, trained on 340 head and neck IMRT plans, generated dose predictions.
• Predictions were used to create IMRT, VMAT, and tomotherapy plans via fallback functionality.
• MCO optimization was employed using predicted doses as constraints, with plan quality assessed against clinical goals.
• Delivery quality assurance (QA) was performed on a subset of plans.

Main Results:

• Dose predictions were accurately replicated across modalities, with minor deviations in critical structures.
• MCO optimization significantly reduced organ-at-risk doses while maintaining target coverage.
• All generated plans demonstrated clinical deliverability, with gamma analysis passing rates exceeding 98%.

Conclusions:

• A model trained on IMRT data can be effectively applied to other radiation therapy modalities.
• Using predictions as MCO constraints offers a flexible warm-start for automated planning.
• These methods show promise for decreasing plan turnaround time and quality variance across different healthcare settings.