Prediction of lung tumor motion using nonlinear autoregressive model with exogenous input

Abstract

The present note addresses the development of a lung tumor position predictor to be used in dynamic tumor tracking radiotherapy, abbreviated as DTT-RT. As there exists 50–500 ms positioning lag in the control of the multi-leaf collimator (MLC) of commercial medical linear accelerators, prediction of future lung tumor position with sufficiently long prediction horizon is inevitable for the successful implementation of DTT-RT. The present article proposes a lung tumor position predictor, which is classified as a nonlinear autoregressive model with exogenous input (NARX). The proposed predictor was trained using seven lung tumor motion trajectories of patients who underwent respiratory gated radiotherapy at Yamaguchi University Hospital. We considered three different prediction horizons, 600 ms, 800 ms and 1 s, which were sufficiently long to compensate for the possible positioning control lag of the MLC. A patient-specific model corresponding to an intended prediction horizon was obtained by training it using the selected tumor motion trajectory with the specified horizon. Accordingly, we obtained three NARX predictors for a single patient. We calculated two performance metrics: the RMS prediction errors and the rate of coverage of the entire tumor trajectory defined by the number of samples of the measured tumor position which was inside the 4 mm cube centered at the corresponding predicted tumor position. The latter quantifies the feasibility of the predictors to generate future gating cubes in the implementation of DTT-RT. The (mean standard deviation) values of the rates of 600 ms, 800 ms and 1 s prediction horizon calculated using the proposed NARX predictors were %, % and %, respectively.