TU-F-CAMPUS-J-03: Prediction of Respiratory Motion Using State Space Models

Authors


Abstract

Purpose:

The purpose of this study is to predict respiratory motion for a few seconds ahead by use of dynamic linear models. The models describe trends and periodic components of time series of respiratory curves obtained on patient's body.

Methods:

To measure spatial coordinates of multiple points on patient's body during respiratory motion, we used a consumer depth camera (Microsoft Kinect) and obtained depth data via triangulation from infrared random dots patterns.To describe a dynamics of respiratory motion, we selected a symplectic form of a harmonic oscillator. As a filter, we selected a particle filter. Particle filter is a technique for implementing a recursive Bayesian filter by Monte Carlo simulations. The key idea is to represent the required posterior density function by a set of random samples and to compute estimates based on these samples.

Results:

Filtered values were calculated as a mean of a filtered distribution. The prediction values were well correlated with the values to be observed.To validate accuracy of our model, predicted depth values were compared with measured ones. The accuracy of our model was roughly 20 % for 4 seconds ahead in average, while the measurement depth accuracy is within 1 mm.

Conclusion:

Time-series modeling using Bayesian inference technique is useful for prediction of respiratory motion.Although prediction accuracy became worse along with the length of forecasting time, we can conclude this method is a promising tool for prediction of patient's motion.

This work was partly supported by the JSPS Core-to-Core Program (No. 23003).

Ancillary