Performance comparison of prediction filters for respiratory motion tracking in radiotherapy

In precision radiotherapy, the intrafractional motion causes substantial uncertainty. Traditionally, the target volume is expanded to cover the tumor in all positions. Alternative approaches are gating and adaptive tracking, which require a time delay as small as possible between the actual tumor motion and the reaction to effectively compensate the motion. Current treatment machines often exhibit large time delays. Prediction filters offer a promising means to mitigate these time delays by predicting the future respiratory motion.

A total of 18 prediction filters were implemented and their hyperparameters optimized for various time delays and noise levels. A set of 93 traces were standardized to a sampling frequency of 25 Hz and smoothed using the Fourier transform with a 3 Hz cutoff frequency. The hyperparameter optimization was carried out with ten traces, and the optimal hyperparameters were evaluated on the remaining 83 traces.

For smooth traces, the wavelet least mean squares prediction filter and the linear filter reached normalized root mean square errors of below 0.05 for time delays of 160 and 480 ms, respectively. For noisy signals, the performance of the prediction filters deteriorated and led to similar results.

Linear methods for prediction filters are sufficient for respiratory motion signals. Reducing the measurement noise generally improves the performance of the prediction filters investigated in this study, even during breathing irregularities.