Customized Prediction of Respiratory Motion - Prediction and Classification of Respiratory Motion - page 101

Table 5.4 Averaged prediction overshoot (CNN vs. RNN)
Prediction overshoot percentage (CNN/RNN)
38.46
ms

115.38
ms

192.3
ms

269.23
ms

346.15
ms

423.07
ms

500
ms

Class
1

12.20/
34.99

13.98/
45.33

15.90/
43.75

16.74/
42.11

19.88/
36.60

13.77/
31.81

19.91/
41.62

Class
2

10.86/
35.32

15.39/
36.63

14.61/
41.59

17.92/
37.65

14.61/
29.96

21.88/
35.62

17.06/
37.31

Class
3

5.20/
28.54

4.16
/34.15

4.93/
35.17

5.20/
40.13

10.36/
30.37

15.13/
32.26

15.14/
38.16

Class
4

3.35/
38.04

4.11/
37.54

4.90/
40.76

9.03/
40.97

9.56/
39.96

11.19/
39.21

15.53/
37.86

Class
5

6.71/
34.45

7.24/
34.10

6.72/
32.31

7.24/
35.41

7.28/
35.22

10.34/
31.93

10.69/
36.93

(Unit: # Overshoot frame/# Total frame: %)

Table 5.5 Comparisons on computational complexity

Methods C-NN R-NN

CPU time used (Unit: Millisecond/#Total frame) 15.11 14.80

5.4.6 Comparisons on Computational Complexity

In this section, we would like to evaluate the computational complexity of the
proposed method. For the comparisons of the computational complexity, we cal-
culate the CPU time used for prediction process over all the total frames.
Table 5.5 shows the average CPU time used for computational complexity over
all the patients. The proposed method needs more computational time for the
prediction process because it is working with three independent RMLPs for each
marker, whereas RNN operates with single target datasets. Moreover, the proposed
CNN has a coupling matrix to organize three independent processes for each
marker. Even though the proposed CNN required more computational time, the
prediction accuracy should compensate for the computational complexity. With
enough computer power these days, the computer time will probably be reduced to
RNN levels within 2 years. We set the prediction time horizon in this study from
38.46 to 500 ms so that any motion can happen within 15 ms on average for the
improved prediction.

5.5 Summary

In this section, we proposed a respiratory motion prediction for multiple patient
interactions using EKF for RNN. When the breathing patterns for the multiple
patients are available, all the patients can be classified into several classes based on