Page 210 - Vibrational Spectroscopic Imaging for Biomedical Applications

186 Cha pte r S i x

were retained for the model. A two-sided t-test was used to select those
PCs that showed the highest significance in discriminating the different
tissue classes. These were used as input for the LDA model. The research-
ers found that contributions of bone to the spectra influenced the classi-
fication model. In a second model, bone signal-corrected PC scores were
used as input to the LDA model and yielded slightly better classification
results. An advantage of using multivariate statistical analysis classifica-
tion models is that they become fairly insensitive to noise, which is
important for in vivo measurements when the acquisition time must be
kept short. However, the number of parameters that can be used as input
into the model must be limited to prevent overfitting of the model.
87
Stone et al. also used LDA in their experiments. Like the previ-
ously described work, PCA was used to reduce the number of vari-
ables in the analysis. Analysis of variance (ANOVA) was utilized to
identify the diagnostically PCs of the spectra. The PCs, used as the
variables for entry into an LDA model, maximized the variance between
groups and minimized the variance within groups. The classification
models were able to separate different types of larynx, bladder, and
esophageal tissues; when tested by cross validation, the models
exhibited high sensitivities and specificities. Interestingly, the
authors noted that instrument variability can decrease the effec-
tiveness of a classification model.
NIR Raman spectroscopy was able to differentiate accurately
between parathyroid adenomas and hyperplasia using a PCA fed
88
LDA diagnostic algorithm. Detection sensitivity for parathyroid
adenomas was 95 and 93 percent for hyperplasia, indicating
Raman spectroscopy is an excellent tool to differentiate between
the two types of parathyroid tissue.
In other work, PCA and LDA were applied to Raman spectra of
a human osteosarcoma cell line (MG63) in order to classify cells
according to the cell cycle phase (G /G , S, and G /M phases). 89
0 1 2
The classification model showed a sensitivity of 93, 67, 60 percent
for cells in the cell cycle block of G /G , S, and G /M respectively,
0 1 2
and specificity of 87, 80, and 93 percent, respectively, as assessed
by a leave-one-out cross-validation.
Chan et al. used PCA and LDA in their experiment that com-
90
bined optical trapping with Raman spectroscopy. In optical trap-
ping Raman spectroscopy, a tightly focused laser beam functions
as both laser tweezers to immobilize a freely diffusing cell and the
excitation source for generating Raman spectra of the trapped cell.
A PCA-LDA classification model was able to discriminate patient-
derived leukemia cells from normal human lymphocytes.
Teh et al. used a PCA-LDA classification approach with Raman
spectra to identify dysplasia from normal gastric mucosa tissue. 91
The model yielded the diagnostic sensitivity of 95.2 percent and
specificity of 90.9 percent for separating dysplasia from normal
gastric tissue.