Real-time prediction and gating of respiratory motion using an extended Kalman filter and Gaussian process regression

Abstract

Motion-adaptive radiotherapy aims to deliver a conformal dose to the target tumour with minimal normal tissue exposure by compensating for tumour motion in real time. The prediction as well as the gating of respiratory motion have received much attention over the last two decades for reducing the targeting error of the treatment beam due to respiratory motion. In this article, we present a real-time algorithm for predicting and gating respiratory motion that utilizes a model-based and a model-free Bayesian framework by combining them in a cascade structure. The algorithm, named EKF-GPR(+), implements a gating function without pre-specifying a particular region of the patient's breathing cycle. The algorithm first employs an extended Kalman filter (LCM-EKF) to predict the respiratory motion and then uses a model-free Gaussian process regression (GPR) to correct the error of the LCM-EKF prediction. The GPR is a non-parametric Bayesian algorithm that yields predictive variance under Gaussian assumptions. The EKF-GPR(+) algorithm utilizes the predictive variance from the GPR component to capture the uncertainty in the LCM-EKF prediction error and systematically identify breathing points with a higher probability of large prediction error in advance. This identification allows us to pause the treatment beam over such instances. EKF-GPR(+) implements the gating function by using simple calculations based on the predictive variance with no additional detection mechanism. A sparse approximation of the GPR algorithm is employed to realize EKF-GPR(+) in real time. Extensive numerical experiments are performed based on a large database of 304 respiratory motion traces to evaluate EKF-GPR(+). The experimental results show that the EKF-GPR(+) algorithm effectively reduces the prediction error in a root-mean-square (RMS) sense by employing the gating function, albeit at the cost of a reduced duty cycle. As an example, EKF-GPR(+) reduces the patient-wise RMS error to 37%, 39% and 42% in percent ratios relative to no prediction for a duty cycle of 80% at lookahead lengths of 192 ms, 384 ms and 576 ms, respectively. The experiments also confirm that EKF-GPR(+) controls the duty cycle with reasonable accuracy.

Personalized Prediction of Glaucoma Progression Under Different Target Intraocular Pressure Levels Using Filtered Forecasting Methods

“…This technology has been critical to aerospace engineering, including guiding Apollo missions to the moon 1 . More recently, it has been used to forecast the development or worsening of chronic diseases 4,8–12 , and we have used it to forecast OAG progression 13 .…”

Machine learning applications in radiation oncology

“…Machine learning techniques are particularly well suited to this application due to the similarity of future breath characteristics with previously recorded breaths. A large body of work has shown neural network, SVM, manifold learning and kernel density estimation can efficiently predict respiratory motion based on previously measured motion traces [153][154][155][156][157][158][159]. Other applications of machine learning involve prediction of motion extent based on tumour size and location in the lungs, automatic diaphragm motion trajectory assessment and incorporation of lung tumour motion into patient setup and prediction of tumour baseline shifts in the short term (approximately 5 s) [160][161][162][163].…”

A review of artificial intelligence applications for motion tracking in radiotherapy

“…Several studies have used ML algorithms for predicting tumour motion based on past motion, 72–89 including in MRI‐guided radiotherapy 90–92 and ultrasound‐guided radiotherapy 93 . A comparison study of ML algorithms was made by Sharp et al.…”

“…Several studies have used ML algorithms for predicting tumour motion based on past motion, [72][73][74][75][76][77][78][79][80][81][82][83][84][85][86][87][88][89] including in MRI-guided radiotherapy [90][91][92] and ultrasound-guided radiotherapy. 93 A comparison study of ML algorithms was made by Sharp et al and showed that most ML algorithms have a lower localisation error compared to no prediction.…”