Models for predicting objective function weights in prostate cancer IMRT.

Area of Science:

• Radiation Oncology
• Medical Physics
• Machine Learning

Background:

• Intensity-modulated radiation therapy (IMRT) for prostate cancer requires precise optimization of objective function weights.
• Patient-specific geometry significantly influences optimal weight selection, posing a challenge for generalized treatment planning.

Purpose of the Study:

• To develop and evaluate advanced machine learning models for predicting multiple IMRT objective function weights.
• To assess the clinical applicability of these models in prostate cancer treatment planning.

Main Methods:

• Retrospective analysis of 315 prostate cancer patients using an inverse optimization method.
• Utilized patient geometry features like overlap volume ratios (OV) and overlap volume histogram slopes (OVSR/OVSB) as predictors.
• Trained and compared logistic regression (LR), multinomial logistic regression (MLR), and weighted K-nearest neighbor (KNN) models against optimal weights.

Main Results:

• Overlap volume (OV) and OVSR were key predictive features for optimal weights.
• LR, MLR, and KNN models demonstrated comparable performance, outperforming population average weights across multiple clinical metrics.
• LR model achieved significant relative improvements in predicted bladder and rectum weights and resulted in treatment plans closer to clinical goals.

Conclusions:

• Machine learning models, including LR, MLR, and KNN, can effectively predict multiple IMRT objective function weights.
• These predictive models yield clinically relevant treatment plans by optimizing trade-offs for sparing multiple organs at risk (OARs).