Review: Prediction of Respiratory Motion - Prediction and Classification of Respiratory Motion - page 29

Table 2.4 Hybrid prediction algorithms of respiratory motion

Methods Prediction error and evaluation

metrics

Features (system)

Adaptive neuro-fuzzy
inference system [ 80 ]

0.628 mm (coached), 1.798 mm
(non-coached), RMSE

25 Hz sample frequency (RPM)

Adaptive tumor tracking
system [ 68 ]

0.8 mm (x-max), 1.0 mm
(y-max), standard deviation

Megavoltage imaging with
infrared system (ELEKTA)

Interacting multiple model
filter [ 81 ]

0.98 mm with 200 ms latency for
5 Hz, RMSE

Kalman CV and CA, Markovian
transition (RPM)

Adaptive motion model
[ 100 ]

1.0-2.8 mm Standard deviation

Hybrid extended Kalman
filter [ 15 ]

Less than 0.15 with 200 ms
latency, normalized RMSE

26 Hz sample frequency
(CyberKnife)

The recurrent network is expressed by the network nonlinearity function b( , , )
with input vectors u(t), the internal state of the recurrent network activities v(t),
and the weight state vector ˆ (t|t - 1). The innovation process a(t) of EKF is
expressed as follows:

a ð t Þ¼ d ð t Þ b ð w ð t j t 1 Þ; v ð t Þ; u ð t ÞÞ: ð 2 : 17 Þ

where b( , , ) is the network nonlinear function of vector-value measurement. The
weight state vector is updated with the Kalman gain G(t) and the innovation
process [ 102 ].
Puskorius et al. proposed a Decoupled EKF (DEKF) as a practical solution for the
computational resource management of covariance value with EKF for RNN [ 101 ].
Suk et al. applied DEKF to the prediction of respiratory motion. They evaluated that
the prediction accuracy of the proposed HEKF and DEKF were less than 0.15 and
0.18 (nRMSE) with 200 ms latency, respectively. They also validated that HEKF
can improve the average prediction overshoot more than 60 %, compared with
DEKF. This method comprehensively organized the multiple breathing signals with
adapting the coupling technique to compensate the computational accuracy,
whereas the computational requirements were increased to improve the prediction
accuracy [ 15 ]. We summarized the prediction accuracy and a representative feature
for each method of the hybrid approach, as shown in Table 2.4 .

2.4 Open Questions for Prediction of Respiratory Motion

Variable open questions on the prediction of respiratory motion are still remained
to be solved in a foreseeable future. In this chapter, we will point out general open
questions for the advanced radiotherapy technology, but open issues are not lim-
ited to the following issues described in this study.