Page 208 - Vibrational Spectroscopic Imaging for Biomedical Applications

184 Cha pte r S i x

from the spectra in the data set to act as the initial cluster center;
(3) all the spectra in the data set are then compared with the clus-
ter centers and assigned to the center that they most resemble;
(4) after all spectra are assigned to a cluster, new cluster centers
are calculated by averaging all the spectra assigned to the cluster;
and (5) the procedure is repeated until a stable solution is reached.
After KCA, a different color was assigned to each cluster and each
grid element of the Raman map was assigned the color of the clus-
ter to which it belonged. The pseudo-color Raman maps were com-
pared with hematoxylin and eosin (H&E)-stained sections and closely
corresponded with the H&E-stained adjacent section. In all the maps,
BCC was separated from its surrounding nontumorous tissue by KCA.
In addition, KCA formed separate clusters for dermis in the vicinity
of the tumor; dermis further away from the tumor; and a dense
inflammatory infiltrate. The difference in the dermal tissue is
most likely due to a decrease in collagen in the dermis adjacent
to tumors.
Nijssen et al. also created a multivariate statistical classifica-
tion model based on the cluster means and the histopathological
78
classification of the cluster means. The classification algorithm
was based on the method of logistic regression, which is a varia-
tion of ordinary regression, useful when the observed outcome is
restricted to two values, usually representing the occurrence or
nonoccurrence of some outcome event. The result was a two-step
classification model where in the first level, dermis is separated
from BCC or epidermis, and then in the second level, BCC is sep-
arated from epidermis. The performance of the model was tested
by a leave-one-out cross validation. The model correctly predicted
BCC from dermis, and showed 100 percent and 93 percent sensitivity
and specificity, respectively.
Thakur et al. used PCA and discriminant functional analysis
(DFA) to evaluate the Raman spectral signatures of mouse mam-
79
mary tissue and associated lymph nodes. Spectra of tumors and
mastitis separated from normal mammary gland tissue and other
pathological states.
PCA has been used with other multivariate statistical tech-
nique such as classical least squares (CLS) and least-squares
analysis (LSA) for the evaluation of pharmaceutical treatment of
80
81
cells and osteoblast differentiation, respectively; in addition
to a quadratic discriminant analysis for the discrimination of cell
response to chemical and physical inactivation. 82
6.6.2 Linear Discriminant Analysis
Linear discriminant analysis (LDA) has proven to be a robust multi-
variate statistical analytical method. Many studies based on Raman
spectra use LDA to create classification models to distinguish between
different types of tissue or cells. LDA is used to find the linear