Review: Prediction of Respiratory Motion - Prediction and Classification of Respiratory Motion - page 25

Update linear model

Infrared camera
- Indirect signal
(~20 frames/s)

Linear adaptive filter
with delay operator
B(q)

Prediction of the
respiratory signal

Nonlinear time series
analysis

Prediction of
tumor motion

Portal images
- Direct signal
(~2 frames/s)

Linear adaptive filter
with delay operator
B(q)

Identify direct
tumor position

Update linear model

Fig. 2.14 Adaptive tumor tracking system with two independent signals. The tumor position is
directly visualized and located by the acquired portal image (direct signal) using a tumor tracking
algorithm without internal fiducial markers [ 68 , 99 ]. Infrared camera signals (indirect signal) are
used to predict respiratory signals using the adaptive filter, and these respiratory signals are
correlated with the portal image to predict the tumor motion [ 98 ]

2b i ;

1
1 þ I c i
a i

l U ð I; a i ; b i ; c i Þ¼ A ; B 2 U ; x ; y 2 I ; i ¼ 1 ; 2 ;

ð 2 : 14 Þ

where I (x and y) are incoming elements, and three parameters (a i , b i , c i ) (referred
to as premise parameters) are continuously updated by training samples using a
gradient descent method [ 80 , 97 ]. Each node in the second layer is a fixed node,
characterized by the product (P) of all the incoming signals, such as w i ¼
l Ai ðÞ l Bi ðÞ; i ¼ 1 ; 2 : Each node in the third layer is a fixed node, characterized
by the normalized ratio (N), such as w i ¼ w i = w 1 þ w ð Þ; i ¼ 1 ; 2 : Each node in
the fourth layer is an adaptive node with a node function, such as w i f i ¼
p i x þ q i y þ r ð Þ; i ¼ 1 ; 2 ; where the parameter set (p i , q i , r i ) (referred to as con-
sequent parameters) are trained by a least squares method. The single node in the
last layer calculates the overall output by aggregating all incoming signals, such as
f ¼ R i w i f i ; i ¼ 1 ; 2 : Kakar et al. validated that the prediction accuracy (RMSE) of
respiratory motion for breast cancer patients was 0.628 mm with coached
breathing and 1.798 mm with free breathing. This method required simpler and
fewer remodeling decorations to implement its nonlinear ability in comparison to
neural networks. However, for other conditions, exemplified by lung patients and
respiration monitoring using spirometry or abdominal straps, it should associate
the breathing signal with the target motion [ 80 ].