Embedding machine learning based toxicity models within radiotherapy treatment plan optimization.

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Biology

Background:

• Radiotherapy (RT) poses challenges due to radiation-induced toxicity (RIT).
• Personalized treatment planning is crucial for minimizing adverse side effects.
• Existing methods often lack integration between toxicity prediction and treatment optimization.

Purpose of the Study:

• To develop and evaluate a personalized radiotherapy treatment planning framework using optimization with constraint learning (OCL).
• To leverage patient-specific data and machine learning to predict and mitigate radiation pneumonitis (RP2+) in non-small cell lung cancer (NSCLC) patients.
• To reduce RIT while maintaining target coverage in RT.

Main Methods:

• Implemented a three-step OCL framework: baseline plan optimization, ML-based RIT prediction, and patient-specific constraint adaptation.
• Utilized classification trees, ensemble methods, and neural networks to predict RP2+ probability.
• Assessed the methodology on four high-risk NSCLC patients, comparing OCL-enhanced plans with conventional plans.

Main Results:

• The OCL framework successfully integrated predictive models into RT planning.
• Mean lung dose and V20 were reduced, leading to an average RP2+ risk reduction from 95% to 42%.
• Tumor coverage was maintained, with minor increases in spinal cord max-dose in some cases.

Conclusions:

• The OCL framework effectively reduces radiation pneumonitis risk in NSCLC patients.
• Integrating patient-specific data into learned constraints enhances personalized RT decision-making.
• This approach bridges the gap between toxicity prediction and treatment optimization for improved patient outcomes.