Customized Prediction of Respiratory Motion - Prediction and Classification of Respiratory Motion - page 89

for tumor tracking techniques [ 1 , 8 , 9 ]. If the acquisition of tumor position and the
repositioning of the radiation beam are not well synchronized, a large volume of
healthy tissue may be irradiated unnecessarily and tumor may be underdosed
[ 10 , 11 ]. Due to the latency, for real-time tumor tracking, the tumor position
should be predicted in advance, so that the radiation beams can be adjusted
accordingly to the predicted target position during treatment [ 1 , 12 ]. Therefore, we
propose a prediction method for respiratory motion to compensate for uncertainty
in respiratory patterns with the correlation of patients breathing datasets.
A number of prediction methods for respiratory motion have been investigated
based on surrogate markers and tomographic images [ 6 , 12 - 26 ]. The previous
methods can be further categorized into two approaches: (1) those that are ''model-
based'', which use a specific biomechanical or mathematical model for respiratory
motion functions or models [ 13 , 14 , 21 , 23 , 24 ]; and (2) those that are ''model-free''
heuristic learning algorithms that are trained based on the observed respiratory
patterns [ 18 , 25 , 27 ]. Generally, model-based methods include linear approaches and
Kalman filter variables that are widely used for the fundamental prediction of
respiratory motion among a variety of investigated methods [ 13 , 14 , 24 ].
A potential drawback of model-based approaches is their inability to learn
highly irregular breathing patterns from training samples [ 27 ]. For accurate pre-
diction of respiratory motion, the breathing pattern information should apply the
respiratory motion prediction to improve prediction accuracy [ 28 ]. Based on
previous studies, the model-free heuristic learning algorithm can be a key
approach for prediction; but, it needs a correction method to compensate for
irregular breathing signals that characterized a variety of breathing patterns.
Accordingly, we have pursued the use of heuristic algorithms to develop system
adaptive loops that have the most general approach, i.e., neural networks (NN).
The contribution of this study is to adopt a clustering method for multiple
patients to get more practical breathing pattern information and to find an accurate
prediction process for an individual class. For the clustering based on breathing
patterns, we present the feature selection metrics. With each feature metric, we can
define a variety of feature combinations and select an optimal feature combination,
i.e., dominant feature selection (Î), and then we can select the appropriate class
number (ˆ) for the analysis of breathing patterns of multiple patients. Finally, we
can predict the respiratory motion based on multiple patient interactions, i.e.,
class-based respiratory motion prediction using interactive degree and neuron
number selection of RNN.
5.2 Prediction Process for Each Patient
For the respiratory motion prediction, we propose to use a supervised-training
feedback system as shown in Fig. 5.1 . The computational complexity of the EKF
depends on the requirement capacity to store and update the filtering-error
covariance matrix. If an RNN has p output nodes and s weights, the asymptotic