Page 302 - Vibrational Spectroscopic Imaging for Biomedical Applications

276 Cha pte r Ni ne

that in Hadamard transform imaging, as is the case in global illumina-
tion with a scanning filter, the generated Raman photons are not all
detected and many are rejected by the mask or the filter, so these tech-
niques can never be as efficient as the point or line imaging techniques
and are only attractive if a high spatial resolution is required.

9.4 Data Analysis: Spectra to Image(s)
From a multidimensional spectral imaging dataset like a Raman map,
one can generate many images, depending on which part of the spectral
information is used for generating the contrast in the image. The sim-
plest images use only the (scaled) intensities of a single wavenumber
channel or a wavenumber band as a gray value (band imaging). One can
extend this using the ratio of two Raman bands as gray value, or com-
bine three Raman bands (or band ratios) to generate a RGB (false or
pseudo) color image (band ratio imaging). In order to make use of all the
information in the spectra at once, more powerful spectral data analysis
methods, based on multivariate statistical techniques, can be applied.
For large images, multivariate analyses can become challenging in
terms of computer power and analysis time because of the amounts of
data that need to be processed. Because the information within a spec-
trum is correlated (intensities at different wavenumbers are not indepen-
dent of each other), and because the spectra within a Raman map often
have a high correlation with each other, it is in many cases useful to
apply a data reduction and orthogonalization technique like principal
30
components analysis (PCA). PCA finds orthogonal directions (princi-
pal components or PCs) in the spectral data space that correspond with
the major sources of variation within that dataset. The amount of varia-
tion captured in each PC diminishes with increasing PC number. Often
the last PCs only represent noise components and can be omitted, which
can give a significant data reduction and noise filtering. Each individual
spectrum can be expressed as a set of scores on these PCs. The scores
of all spectra on a certain PC can again be used as a gray value to
generate an image, now using information from the whole spectrum
to generate image contrast. With the scores of the first three PCs,
being the PCs that represent the major sources of variation, one can
generate an RGB image that has the highest spectral contrast infor-
mation density possible, but this need not always be the most infor-
mative contrast possible.
Multivariate spectral data analysis techniques can be separated in
two categories: classification techniques to group spectra on basis of spec-
tral similarities and quantification techniques for (bio)chemical composi-
tion analyses. For each, two subcategories are distinguished: supervised
and unsupervised. Supervised analysis makes use of models built on pre-
viously measured data, where the unsupervised models only use an
algorithm working on the current dataset. Below, the basic principles of
the currently used methods are described. For more extensive reviews