Publication

Abstract : Purpose This study investigates whether machine learning models trained exclusively on RUBY phantom shift data from CBCT and kV/MV imaging can accurately predict daily patient setup shifts in Head & Neck radiotherapy. Methods A total of 12,600 matched imaging fractions from Head & Neck treatments were analysed, each comprising six input features (shifts in X, Y, Z from CBCT and kV/MV imaging) and three output variables (X, Y, Z patient shifts relative to planning CT). Eight regression models were evaluated: Random Forest, AdaBoost, Gradient Boosting, XGBoost, LightGBM, K-Nearest Neighbors, Support Vector Regressor, and Multilayer Perceptron. Performance metrics included mean absolute error (MAE), root mean square error (RMSE),coefficient of determination (R2),and prediction accuracy within ±0.5 mm and ±1 mm. Results The Random Forest, XGBoost, and LightGBM models achieved the best performance, each with MAE ≈ 0.254 mm, RMSE ≈ 0.510 mm, and R2 ≈ 0.79 across all axes. These models also achieved mean accuracies of 87.9 % within ±0.5 mm and 93.6 % within ±1 mm. Ensemble tree-based methods outperformed other approaches, with AdaBoost showing the lowest accuracy. Feature importance analysis identified CBCT Y and Z phantom shifts as the strongest predictors, reflecting the superior geometric fidelity of volumetric imaging in capturing translational deviations. CBCT-derived features contributed the majority of predictive power, particularly in lateral and longitudinal directions, while kV/MV shifts had relatively lower influence. Conclusion Phantom-based machine learning models can accurately predict daily setup deviations in Head & Neck radiotherapy with submillimetre precision. This approach could enhance adaptive workflows, support selective imaging protocols, and enable automated pre-treatment verification, thereby improving both treatment efficiency and patient safety.