Purpose: Intra-fractional tumor motion prediction is still an important technical challenge for achieving accurate real-time adaptive radiation therapy (RT-ART). Indeed, effective prediction methods for RT-ART of moving prostate tumors have not been developed yet because of the complicated motion. The purpose of this study is to test the predictability of prostate tumor motion.
Methods: A statistical modeling approach was used for predicting four datasets of the prostate motion time series. The motion has complex and irregular changes with time, while it often contains a trend component, i.e., long-term change in the level of the time series. It is thus expected that modeling the trend can provide a rough prediction of the motion. As a simple approach to predict the trend, this study adopted an autoregressive (AR) model-based method with the differencing of the time series. The differencing was performed for making the time series stationary and capturing the trend. An AR model was then trained to predict the trend component. The prostate motion prediction was reconstructed by cumulatively summing the AR model output. The AR model-based prediction was evaluated on the four datasets and compared with constant position (CP) model as a conventional method and constant velocity (CV) model as a benchmark method.
Results: On average, the prediction errors of the AR, CP, and CV models were 0.27 mm, 0.37 mm, and 0.40 mm, respectively. The AR model achieved a 35% error reduction compared to the CP model while the CV model didn’t outperform the CP model.
Conclusion: The AR model-based approach first demonstrated that the prostate tumor motion is partially predictable. Although the current prediction performance is yet far from clinical use, further prediction error reduction with more appropriate modeling potentially provides a substantial improvement in the RT-ART accuracy for prostate cancer.
Modeling, Organ Motion, Prostate Therapy
TH- External Beam- Photons: Motion management - intrafraction