Comparison between acceleration-enhanced adaptive filters and neural network filters for respiratory motion prediction

Abstract

The prediction of respiration-induced organ motion is crucial in some applications such as dynamic delivery of radiation dose. In this paper, we have proposed the novel approach to construct an acceleration-enhanced (AE) filter that is comprised of two independent adaptive channels. The filters use the adapted position and adapted acceleration, together with a weight factor to provide prediction for respiratory motion. The proposed AE approach is universal and can be applied to the different filters. The performances of the adaptive normalized least mean square (nLMS) filter, the artificial neural network (ANN) filter, and their AE counterparts were compared for respiratory motion prediction during normal and irregular respiration. The results revealed that the adaptive ANN and nLMS filters were successful to perform predictions for normal and irregular respiration, respectively. AE filters showed more accurate prediction than their conventional counterparts. Implementing the AE approach, it was observed that the AE-ANN filter had the best performance in the prediction of normal respiratory motion, whereas the AE-nLMS filter excelled in the prediction of irregular respiratory motion.