Customized Prediction of Respiratory Motion - Prediction and Classification of Respiratory Motion - page 100

Fig. 5.7 b, the figure shows that most estimation values of the proposed CNN are
within the upper and lower bounds, whereas some estimation values of RNN lie
outside the confidence level. The time index duration out of the overshoot mar-
ginal value (c) for this particular patient is 23.22 % using CNN and 46.45 % using
RNN, respectively.
For the prediction overshoot comparison of all 130 patients, we calculated the
number of the total frame and overshoot frames for each patient and show the
prediction overshoot frames for the proposed CNN and the alternate RNN with
respect to all the classes in Fig. 5.8 . Figure 5.8 shows that most of the prediction
overshoot numbers for CNN are much smaller than those for RNN over all the
patients, even though there are some exceptions, i.e. five patients in the class 1
(patient numbers 8, 10, 12, 47 and 86), and three patients in the class 2 (patient
numbers 20, 22 and 38). For five classes among the 130 patients, we calculate the
averaged overshoot frames over the total frame with respect to the prediction time
horizon as shown in Table 5.4 . As shown in prediction overshoot of Table 5.4 , the
proposed CNN does not directly address the criterion of overshoot regarding the
class selection among multiple patients; therefore the larger size of patients may
have relatively large overshoot for in the particular class.
The averaged overshoot frames of RNN are more than 35 % overall the classes,
whereas the averaged overshoot frames of the proposed CNN are within 13 % in
the short time prediction. Note that averaged overshoot frames are less than 7 % in
the short and middle time prediction of classes 3, 4 and 5. Based on Table 5.4 , the
proposed CNN shows more reliable prediction in comparison with the alternate
RNN over all the patients.

CNN RNN

100

50

0

5 10 15 20 25 30 35 40

Class 1 (40 Patients)

100

50

0

5 10 15 20 25

Class 2 (27 Patients)

100

50

0

2 4 6 8 10 12

Class 3 (13 Patients)

100

50

0

5 10 15 20 25

Class 4 (29 Patients)

100

50

0

2 4 6 8 10 12 14 16 18 20

Class 5 (21 Patients)

Fig. 5.8 Prediction overshoot comparison over all the patients with 192 ms latency. The
prediction overshoots with CNN were improved in the most of the patients except five patients in
the class 1 (patient numbers 8, 10, 12, 47 and 86), and three patients in the class 2 (patient
numbers 20, 22 and 38)