Tumor tracking with non-linear internal/external correlation models in the presence of respiratory motion baseline drifts and phase shifts

Purpose In CyberKnife® respiratory tracking, tumor positions are predicted from external marker positions using correlation models. With available models, prediction accuracy may deteriorate when respiratory motion baseline drifts occur. Previous investigations have demonstrated that for linear models this can be mitigated by adding a time-dependent term. In this study, we have focused on added value of time-dependent terms for the available non-linear correlation models, and on phase shifts between internal and external motion tracks. Methods Treatment simulations for tracking with and without time-dependent terms were performed using computer generated respiratory motion tracks for 60.000 patients with variable baseline drifts and phase shifts. The protocol for acquisition of X-ray images was always the same. Tumor position prediction accuracies in simulated treatments were largely based on cumulative error-time histograms and quantified with R95: in 95% of time the prediction error is