The comparative performance of four respiratory motion predictors for real-time tumour tracking

Prediction of respiratory motion is essential for real-time tracking of lung or liver tumours in radiotherapy to compensate for system latencies. This study compares the performance of respiratory motion prediction based on linear regression (LR), neural networks (NN), kernel density estimation (KDE) and support vector regression (SVR) for various sampling rates and system latencies ranging from 0.2 to 0.6 s. Root-mean-squared prediction errors are evaluated on 12 3D lung tumour motion traces acquired at 30 Hz during radiotherapy treatments. The effect of stationary predictor training versus continuous predictor retraining as well as full 3D motion processing versus independent coordinate-wise motion processing is investigated. Model parameter optimization is performed through a grid search in the model parameter space for each predictor and all considered latencies, sampling rates, training schemes and 3D data-processing modes. Comparison of the predictors is performed in the clinically applicable setting of patient-independent model parameters. The considered predictors roughly halve the prediction errors compared to using no prediction. When averaging over all sampling rates and latencies, prediction errors normalized to errors of using no prediction of 0.44, 0.46, 0.49 and 0.55 for NN, SVR, LR and KDE are observed. The small differences between the predictors emphasize the relative importance of adequate model parameter optimization compared to the actual prediction model selection. Thorough model parameter tuning is therefore essential for fair predictor comparisons.