Raman Spectroscopy for Early Cancer Detection, Diagnosis and Elucidation of Disease-Specific Biochemical Changes - Emerging Raman Applications and Techniques in Biomedical and Pharmaceutical Fields - page 319

Fig. 13.3. Scatter plot of linear discriminant scores for each spectrum in the biopsy
targeting model when tested against all the others, colour coded for consensus
pathology
Study of Model Architecture
Initially a study of the architecture of the PC-fed LDA model was made.
The three-group model used in this study was constructed from the first 25
principal components of the spectral data set. Each sequential principal com-
ponent describes a decreasing level of variance from the mean of the spectral
data set. Therefore the later PCs will tend to describe more subtle spectral
changes across the data set. The balance must be achieved between excluding
subtle potentially diagnostic information and including too much noise from
the higher level PCs.
The principal component contributions to the model have been studied.
In this model the first 25 PCs were included. Figure 13.4 shows the mean
PC scores for each of the three pathology groups (top). This enables visual-
isation of the most significant variations between the groups. The ANOVA
F -value is plotted next (middle) to indicate the PCs with the most statisti-
cally significant differences between the groups. The bottom plot is the weight
of each PC used to construct the LD functions. It can be seen that the first
10-12 PCs are the most significant (middle). However if the LD weights are
studied in isolation, it appears that higher PCs are given more weight in the
model. This misconception is overcome by studying Fig. 13.5, which shows