Page 209 - Vibrational Spectroscopic Imaging for Biomedical Applications

W idefield Raman Imaging of Cells and T issues 185

combination of features which best separate two or more classes.
Unlike PCA, LDA explicitly attempts to model the difference between
the classes of data. Typically, a dimensional reduction technique, such
as PCA, is used to reduce the variables and identify significant differ-
ences in the data (PCs). The identified PCs are then used as input into
the LDA model to help create a classification algorithm which mini-
mizes misclassification.
Combining Raman spectroscopy with optical confocal micros-
copy, called confocal Raman microscopy, adds spatial resolution to
spectral information. Taleb et al. used confocal Raman microscopy to
83
map two different prostatic cell lines, PNT1A and LNCaP. In this
study, different data preprocessing methods, partial least-squares
(PLS) and adjacent band ratios (ABRs) were used as input into LDA
to model and classify the two cell lines. The study demonstrated that
PLS discriminate analysis and ABRs of Raman data as input into LDA
can identify subtle differences between benign (PNT1A) and malig-
nant (LNCaP) prostate cells in vitro.
Mahadevan-Jansen et al. used near-infrared Raman spectroscopy
84
to differentiate cervical precancers from normal tissue. Multivariate
statistical algorithms were developed for tissue differentiation using
a five-step procedure: (1) data processing; (2) dimension reduction
using PCA; (3) selection of diagnostically important PCs using the
unpaired Student’s t-test; (4) development of a classification algo-
rithm using Fisher’s discriminant analysis (FDA); and (5) unbiased
evaluation of algorithm performance using cross validation. FDA,
similar to LDA, is a statistical technique that uses linear combinations
of independent variables to discriminate between two groups such
that the misclassification rates are minimized. The algorithms were
able to differentiate squamous intraepithelial lesions (SIL) from
non-SIL samples.
Similarly, PCA and LDA of Raman spectra of cervical tissue
were used to distinguish between normal cervical tissue, cervical
85
intraepithelial neoplasia (CIN), and invasive carcinoma. Based on
cross-validation results, sensitivity and specificity values were calcu-
lated as 99.5 and 100 percent respectively for normal tissue, 99 and
99.2 percent respectively for CIN, and 98.5 and 99 percent respec-
tively for invasive carcinoma.
LDA was used to create a classification model in another study.
The potential of Raman spectroscopy for in vivo classification of nor-
mal and dysplastic tissue was investigated by Bakker Schut et al. 86
The Raman dispersive spectra were acquired from normal and
dysplastic rat palate tissue in vivo using a fiber optic probe. A proce-
84
dure similar to that employed by Mahadeven-Jansen et al. was
used to create an LDA classification model. PCA was used on the
model set to orthogonalize and reduce the number of parameters
needed to represent the variance in the spectral data set. PCs that
accounted for more than 1 percent of the variance in the data set