Infrared and Raman Spectroscopy and Imaging of Breast Cancer - Vibrational Spectroscopy for Tissue Analysis - page 156

two ANNs: one to classify spectra as either cancer or not cancer, and another
to classify the cancers as either FA or DCIS. This improved the results sub-
stantially, but there was one sample from which most of the spectra were
misclassified, highlighting the important problem of interpatient variability.
Krafft et al. [63] used SIMCA classification of FTIR microspectroscopic
images to identify the primary tumours of brain metastases. They had
20 samples, split 5/15 into training and validation sets. There was one train-
ing sample for each type of primary tumour. They used min-max normalisa-
tion to account for thickness differences. The reclassification of the training
data was >90% for each sample. The classification of the independent sam-
ples was quite good, with all samples but one having >50% of their spectra
correctly classified; the average correct classification percentage was 79%.
Bhargava et al. and Fernandez et al. [58,64] present results of a study on
tissue microarrays , arrays formed by embedding small core biopsies of a sample
in a wax block, which is then microtomed and mounted on a slide (transpar-
ent window, but they mention reflective slides as well). Their samples were of
prostate tissue. They suggest that standard chemometric data processing and
classification-clustering algorithms are too slow for the large datasets pro-
duced in their study. They use a different approach based on Bayesian classi-
fication of spectral metrics. The spectral metrics are things like peak positions,
widths, and heights, and are isolated from representative spectra manually
by a spectroscopist. This appears to be simply a way to reduce the number of
variables while retaining specific, local information. The set of spectral met-
rics is further reduced, algorithmically, to only the useful ones. They defined
the classification task as determining the probability that an unknown spec-
trum originated from a particular class of tissue. Their classification approach
is similar to LDA, but utilises Bayes's theorem to derive discriminant func-
tions automatically. They state that the approach relies heavily on having a
very large dataset available, since the weights in the discriminant functions
are overly determined by the underlying distributions of components.
Bambery et al. [66] studied a malignant glioma tumour in a rat brain via
FTIR imaging. They used reflection-absorbance. The sampling area covered
nearly the entire rat brain section (11 mm × 11 mm; 4-µm thick), and the
tumour was a few milimeters in diameter. Hierarchical clustering was easily
able to identify the tumour region. Comparison was made to a healthy rat
brain and to histological examination of the brain with the tumour.
Summary
To summarise, both the Raman and FTIR spectra are well known for their
sensitivity to composition and three-dimensional structure of biomolecules.
The biochemical changes in the subcellular levels developing in abnormal