Biomedical Engineering Reference
In-Depth Information
To derive B ,
Ω N is defined as
C
CA
Ω N =
(18)
.
CA N 1
and B is obtained by
Ω N Ω N ) 1 (
Ω N y T ),
B
=
(
(19)
where y is the vector of the frequency samples of the backscattered data. Thus, a i
could be derived from Eq. 15. Now that all the parameters of the model in Eq. 3 are
derived, the frequency response of the system could be reconstructed.
Lesion Classifier
Decay rate and oscillation frequency of complex harmonics extracted from
frequency-domain data are determined by α and R in Eq. 3, respectively. As discussed
in the previous section, the decay factors or poles of the harmonics are dependent on
the shape of the scattering point while phase factors R are determined by the distance
of the scatterer. A sample of time-domain backscattered signal and its frequency re-
sponse is shown in Fig. 3 . In this study, we investigate the feasibility of using shape
parameter α to distinguish benign and malignant lesions. In fact, as mentioned be-
fore, benign lesions tend to be more regular and spherical while malignant lesions are
usually irregular and highly spiculated. This will affect the poles in the backscattered
signal. The classifier captures the changes in poles of the frequency-domain signals
from benign and malignant lesions and hence, classifies the lesion type. MATLAB
(matrix laboratory) environment was used to build neural network classifiers. Algo-
rithms written for the processing of the data set and the computation of the results
are thus in MATLAB's native language. In this study, a total of 25 harmonics were
extracted from each response initially which is the number of singular values of the
Hankel matrix. Optimizing the number of the poles will be discussed later in this
chapter.
Neural networks constructed were trained at least 30 times with the initial weights
and biases of the networks re-initialized before each trial. Only the trial with the
highest accuracy among others with the same network structure has been chosen for
discussion to be compared with other networks that likewise have been chosen for
highest accuracy from their respective network structures.
A pattern recognition network, typically a feed-forward back-propagation network
with tan-sigmoid transfer functions in both hidden layers and output layer, is used
in this application.
Search WWH ::




Custom Search