Biomedical Engineering Reference
In-Depth Information
and PET. UWB screening and diagnosis of breast cancer has been investigated by
researchers for over a decade [
2
-
21
]. Low-cost equipments and harmless radiation
make UWB imaging a promising modality in early detection of breast cancer. UWB
pulse could be used for both detection and classification of the breast lesion. The fea-
sibility of using UWB pulse to detect breast lesions is studied in [
2
,
3
,
7
,
10
,
20
,
22
].
In this approach, a breast is illuminated by a UWB pulse and the backscattered sig-
nal is collected by an array of antennae. The UWB beam is synthetically focused
spatially to yield a map of dielectric variations within the breast medium. Breast
lesions normally exhibit irregularly high dielectric constant and thus appear brighter
in the synthetic beamformer images. In addition to localization of the lesion, UWB
could also be used to determine malignancy of the lesion [
4
,
6
,
8
,
9
,
15
,
21
]. In [
15
],
the authors have investigated the feasibility of using UWB pulse in characterizing
the lesion based on its size and shape. In this study, two basis selection methods,
local discriminant basis and principal component analysis, are employed to construct
and evaluate a UWB lesion classifier. The effect of lesion morphology on temporal
response of the backscattered UWB pulse is studied in [
6
]. It is shown that the le-
sion morphology affects the complex natural resonances in the late-time backscatter
response of UWB pulse and hence, the potential use of this method for discrimi-
nating between benign and malignant masses. The use of contrast agent for lesion
classification is studied in [
5
]. In this study, the damping factors of the differentiated
backscatter responses are compared before and after infusion of the contrast agent
to detect anomalies. In [
8
], support vector machine (SVM) classifier is suggested
and compared with linear discriminant analysis (LDA) and quadratic discriminant
analysis (QDA) classifiers. Multi-input multi-output (MIMO) radar techniques are
used in [
4
] to enhance resonance scattering based lesion classification. In [
21
], the
early time portion of the backscattered breast response is passed through a correlator
to quantify the degree of lesion ruggedness and hence, the malignancy of the lesion
is determined.
The main objective of this chapter is to investigate the feasibility of classifying
breast lesions using their frequency-domain UWB backscatter responses. A UWB
breast lesion classifier is introduced based on the frequency-domain signature of
the lesion. Complex harmonics are extracted from the frequency response and com-
bined to form an input vector to a neural network classifier. The classifier is then
trained using a training subgroup of vectors obtained from numerical analysis of the
lesion responses. Different architectures of neural networks and their performances
in characterizing lesions are evaluated and discussed.
This chapter is organized as follows: The state-of-the-art UWB breast lesion
detection and classification is given in “Introduction”. Details on the numerical anal-
ysis of the lesion backscattered signal are given in the section, “Numerical Breast
Model”. “Complex Harmonic Estimation” discusses the frequency-domain process-
ing of the backscatter signal followed by “Lesion Classifier”,which provides the
frequency-domain classifier design and architecture. “Results” evaluates the perfor-
mance of different architectures of neural networks. Finally, “Conclusion” concludes
this chapter.
Search WWH ::
Custom Search