Performance enhancement in respiratory tumor motion prediction

Abstract

Introduction

The breathing motion moves internal organs and targeted regions determined by radiation therapy planning. For the radiation therapy, accurate prediction for breathing motion is of great interest as the outer targeted region treatment could endanger sensitive tissue. In this study, a prediction algorithm with adaptive support vector regression (aSVR) was proposed and compared with the adaptive neural network (ANN) algorithm considering the prediction accuracy and the computational costs of training and predicting.

Methods

Respiration data from 87 patients treated by radiation therapy, were acquired with an optical marker at 30 Hz. Five types of prediction filters with the ANN or aSVR filters, were implemented and their performance was compared according to the size of the sliding window (2.5 and 5.0 sec), and the prediction latencies (100, 200, 300, 400, and 500 msec). Training and testing of the prediction algorithms with aSVR and ANN were performed. The root mean square error (RMSE) was used as the accuracy metric.

Results

The aSVR with a radial basis function (RBF) kernel outperformed other prediction filters, including not only various types of ANN filters but also the aSVR with a linear kernel. A sliding window of 2.5 sec significantly and independently enhanced the overall accuracy. Otherwise, the training and prediction testing times were significantly prolonged in case of aSVR with an RBF kernel.

Conclusions

The aSVR filter with the RBF kernel is in all cases superior to other filters regarding its accuracy

it also shows clinically applicable results from the viewpoint of training and predicting time, which may be effective for predicting patient breathing motion and thus enhancing the efficacy of radiation therapy.