490
S. Biswas et al.
hence their characteristic spectra are similar but show small differences in some
specific features. Chemical concentrations can be measured by recording the spectral absorbance at specific vibrational frequencies, while subtle molecular structural changes are analysed by studying the spectral peak shifts, band shapes, relative
intensity changes portrayed in the absorption spectra of the respective classes [48].
In most IR imaging experiments, determinations are made pixel-by-pixel.
An absorption spectrum is obtained for every pixel of a 2d (thin in terms of
beam interaction in one dimension compared to the other two dimensions), thus
leading to the formation of the so-called “data cube”. the acquired data is usually
analyzed in two ways: either by obtaining chemical images at every wavenumber or
by extracting absorption spectra at every pixel. For applications in histopathology,
both methods have advantages but spatial information in imaging is typically of
more interest and spectral information is usually compressed in some manner to be
spatially appreciated. the most common is the use of color coding
2
, in which a color
coded image demonstrates the presence of multiple cell types in a tissue.
16.3.4.2 Building a Classifier
to relate spectral features to underlying biologic or physiologic assessment, a computational approach typically based on statistical properties of the data is most commonly used. typically, spectral features from the absorption spectra are fed into
pattern recognition algorithms to extract information. An overview of the various
steps that may be typically involved in a specific protocol are shown in Fig. 16.8
[3]. the key elements of building the classifier are discussed below.
Data Acquisition While it is obvious that data acquisition is the first step, a critical
consideration at this stage is to ensure that the data are representative of the full data
set. If this is not ensured, developed methods may be subject to the dual pitfalls of
chance and bias, limiting the applicability of any developed protocol. the data set
to develop protocols must also be sufficiently large for the results to have required
statistical power.
Pre-Processing Pre-processing collectively refers to a set of computational processes that aim to enhance the characteristics of the acquired data, making analysis more accurate. the use of algorithms such as the minimum Noise Fraction
to reduce noise for more accurate analysis, baseline correction to remove effects
due to scattering and removal of water vapor signals are all examples of common
steps. Appropriate pre-processing plays a very important role in both analyzing and
extracting important chemical information from recorded spectra and paves the path
to building accurate, fast and robust classifiers.
Classification the goal of classification methods is to categorize data into subsequent groups, which can provide with powerful information about the molecular
2
Note that some texts refer to this practice as “False Color”. this term is misleading as all colors
are true. “Color-coded” images or “Representative color” images are probably better monikers.
S. Biswas et al.
hence their characteristic spectra are similar but show small differences in some
specific features. Chemical concentrations can be measured by recording the spectral absorbance at specific vibrational frequencies, while subtle molecular structural changes are analysed by studying the spectral peak shifts, band shapes, relative
intensity changes portrayed in the absorption spectra of the respective classes [48].
In most IR imaging experiments, determinations are made pixel-by-pixel.
An absorption spectrum is obtained for every pixel of a 2d (thin in terms of
beam interaction in one dimension compared to the other two dimensions), thus
leading to the formation of the so-called “data cube”. the acquired data is usually
analyzed in two ways: either by obtaining chemical images at every wavenumber or
by extracting absorption spectra at every pixel. For applications in histopathology,
both methods have advantages but spatial information in imaging is typically of
more interest and spectral information is usually compressed in some manner to be
spatially appreciated. the most common is the use of color coding
2
, in which a color
coded image demonstrates the presence of multiple cell types in a tissue.
16.3.4.2 Building a Classifier
to relate spectral features to underlying biologic or physiologic assessment, a computational approach typically based on statistical properties of the data is most commonly used. typically, spectral features from the absorption spectra are fed into
pattern recognition algorithms to extract information. An overview of the various
steps that may be typically involved in a specific protocol are shown in Fig. 16.8
[3]. the key elements of building the classifier are discussed below.
Data Acquisition While it is obvious that data acquisition is the first step, a critical
consideration at this stage is to ensure that the data are representative of the full data
set. If this is not ensured, developed methods may be subject to the dual pitfalls of
chance and bias, limiting the applicability of any developed protocol. the data set
to develop protocols must also be sufficiently large for the results to have required
statistical power.
Pre-Processing Pre-processing collectively refers to a set of computational processes that aim to enhance the characteristics of the acquired data, making analysis more accurate. the use of algorithms such as the minimum Noise Fraction
to reduce noise for more accurate analysis, baseline correction to remove effects
due to scattering and removal of water vapor signals are all examples of common
steps. Appropriate pre-processing plays a very important role in both analyzing and
extracting important chemical information from recorded spectra and paves the path
to building accurate, fast and robust classifiers.
Classification the goal of classification methods is to categorize data into subsequent groups, which can provide with powerful information about the molecular
2
Note that some texts refer to this practice as “False Color”. this term is misleading as all colors
are true. “Color-coded” images or “Representative color” images are probably better monikers.