Infrared and Raman Spectroscopy and Imaging of Breast Cancer - Vibrational Spectroscopy for Tissue Analysis - page 134

acids. Similarly, increase in intensity and shifting of peaks in the area of 1700
to 1500 cm −1 (carbonyl stretching and amide-bending vibrations in the DNA)
confirm the changes in the chemical structure of normal tissue. An overall shift
of peaks in the fingerprint region (1700 to 800 cm −1 ) is observed in the spec-
trum of breast cancer tissue. Figure 5.18 shows the shift in these bands as char-
acterised by an infrared image of the amide I band between 1700-1600 cm −1 .
The deeper the red colour, the greater the infrared shift in this spectral region.
Useful information pertaining to carbonyl stretching and amide-bending
vibrations in the DNA can be extracted. This data can be extremely useful in
defining infrared images that show the shift of amide bands between 1700-
1500 cm −1 . The images extracted show that this shift is significant, sometimes
varying by as much as 20 cm −1 . In addition, principal components analysis
(PCA) can be applied to the images to extract further data producing images
based on the principal components.
Data Handling requirements and Spectral Pretreatments
A single image usually consists of 64 × 64 = 4096 spectra, each consisting of
measurements at several hundred to a few thousand wavelengths. Assuming
1000 wavelengths, there are roughly 4 million data points in one image. This
equates to about 30 MB. Many images may be tiled together to image a larger
sample, and when images from many specimens from many samples are
combined, datasets can be several GB. Computations (particularly hierarchi-
cal clustering) on the dataset as a whole can then be challenging, and data-
reduction methods (such as PCA or clustering) may be applied to individual
images prior to analysing the data set as a whole.
Sometimes, the spectra are analysed without any pretreatments [67]. More
commonly, however, varieties of numerical treatments are used to reduce
the influence of artefacts. First [71,72] and second [68] derivatives are used
to reduce baseline effects and to enhance small features. An alternative
approach to removing baselines is to subtract a linear function directly
[59,63]. Often, some kind of scaling is used as well, such as an absorbance ratio
[59,61,63] (often applied to the amide I band or to the difference between the
minimum and maximum absorbance values) or the standard normal vari-
ate (SNV) transform [60,68], in which each spectrum is scaled to have zero
mean and unit variance. The purpose of scaling is to emphasise differences
in the shapes, rather than intensities, of the spectra, and it is particularly
justified when there are variations in sample thickness due to imperfections
in sample preparation.
Spectra are usually screened for certain quality criteria such as having
acceptable absorbance levels or signal-to-noise ratios. Liquid nitrogen-
cooled FPA cameras are apparently prone to developing pixel defects due
to thermal cycling, and these pixels should be identified and excluded from
the analysis.