Predicting patient setup shifts in daily radiotherapy using machine learning on phantom-based CBCT and kV/MV data.

Area of Science:

• Radiotherapy Physics
• Medical Imaging
• Machine Learning in Healthcare

Background:

• Accurate patient positioning is critical in Head & Neck radiotherapy to ensure dose delivery to the tumor while sparing organs at risk.
• Daily setup variations can compromise treatment accuracy, necessitating robust verification methods.
• Machine learning offers potential for predicting and mitigating these setup errors.

Purpose of the Study:

• To evaluate the efficacy of machine learning models trained on phantom shift data for predicting daily patient setup shifts in Head & Neck radiotherapy.
• To determine if models trained solely on Cone-Beam Computed Tomography (CBCT) and kV/MV imaging phantom data can achieve submillimetre accuracy.
• To identify the most influential imaging features for predicting patient setup deviations.

Main Methods:

• Analysis of 12,600 Head & Neck treatment fractions, using CBCT and kV/MV imaging shifts as input features and patient shifts (X, Y, Z) as output variables.
• Evaluation of eight regression models including Random Forest, XGBoost, and LightGBM, assessing performance via Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²).
• Calculation of prediction accuracy within ±0.5 mm and ±1 mm tolerance levels.

Main Results:

• Ensemble tree-based models, particularly Random Forest, XGBoost, and LightGBM, demonstrated superior performance with MAE ≈ 0.254 mm, RMSE ≈ 0.510 mm, and R² ≈ 0.79.
• These top models achieved mean accuracies of 87.9% within ±0.5 mm and 93.6% within ±1 mm.
• CBCT Y and Z phantom shifts were identified as the strongest predictors, indicating volumetric imaging's importance in capturing translational deviations.

Conclusions:

• Machine learning models trained on phantom data can accurately predict daily setup deviations in Head & Neck radiotherapy with submillimetre precision.
• This predictive capability can significantly enhance adaptive radiotherapy workflows and support selective imaging protocols.
• The approach holds promise for automated pre-treatment verification, improving both treatment efficiency and patient safety.