Development and Clinical Validation of a Protocol-Agnostic Machine Learning Platform for Automated Treatment Planning...

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning

Background:

• Knowledge-based planning (KBP) in external beam radiation therapy traditionally relies on protocol-specific models trained on limited data.
• This approach can limit generalizability and optimization potential across diverse patient populations and treatment sites.
• A need exists for more adaptable and robust KBP strategies.

Purpose of the Study:

• To develop and validate a protocol-agnostic machine learning platform, termed "Predictive Planning," for KBP in external beam radiation therapy.
• To assess the dosimetric outcomes of plans generated using this platform compared to standard clinical plans.

Main Methods:

• Developed general-purpose machine learning models for organ at risk (OAR) dose-volume histogram prediction using over 5,000 retrospective treatment plans from various sites (Head and Neck, Thoracic, Abdominal, Pelvis).
• Deployed models in a commercial system (Plan AI) to replan 72 clinical cases, generating "Predictive Plans" (PPs) based on predicted optimization objectives.
• Compared OAR and planning target volume (PTV) dose metrics between PPs and original clinical plans.

Main Results:

• Predictive Plans (PPs) demonstrated statistically significant reductions in mean dose for numerous OARs across all treatment sites, including Brain, Brainstem, Spinal Cord, Heart, Kidney, Bladder, and Rectum.
• No OAR dose metrics were found to be statistically significantly worse in PPs compared to clinical plans.
• Mean PTV coverage was significantly higher for Head and Neck and Thoracic cases, and equivalent for Abdominal and Pelvis cases.

Conclusions:

• Predictive Planning offers a novel approach to KBP, shifting from protocol-specific to protocol-agnostic, disease-site-specific models trained on large, heterogeneous datasets.
• Plans generated using predicted optimization objectives resulted in equivalent or superior dosimetric outcomes, highlighting the potential of this platform to improve radiation therapy planning.