Lung tumor motion prediction based on multiple time-variant seasonal autoregressive model for tumor following radiotherapy

This paper presents a new lung tumor motion prediction method for tumor following radiation therapy. An essential core of the method is accurate estimation of complex fluctuation of time-variant periodical nature of lung tumor motion. Such estimation can be achieved by using a multiple time-variant seasonal autoregressive integral moving average (TVSARIMA) model in which several windows of different lengths is used to calculate correlation based time-variant period of the motion. The proposed method provides the final predicted value as a combination of those based on different window lengths. We have tested unweighted average, multiple regression, and multi layer perceptron (MLP) for the combination method by using real lung tumor motion data. The proposed methods with multiple regression and MLP based combinations showed high accurate prediction and are superior to the single TVSARIMA based prediction. The most highest prediction accuracy was achieved by using the MLP based combination. The average errors were 0.7953±0.0243[mm] at 0.5[sec] ahead and 0.8581±0.0510[mm] at 1.0[sec] ahead predictions, respectively. The results clearly demonstrate that the proposed method with an appropriate combination of several TVSARIMA is useful for improving the prediction performance.