494
S. Biswas et al.
Unsupervised Learning Common unsupervised learning methods include
K-means and hierarchical clustering analysis. In this type of classification, prior
knowledge of the classes within the data is not available and, indeed, the focus is
often on discovering the underlying structure. the data acquired is divided by the
algorithm used into clusters or groups. this form of classification is highly valuable
as a visualization and discovery tool in IR imaging. the results from unsupervised
methods may be used as is, especially when outliers are to be detected or a ready
reference (such as a corresponding stained tissue section) is available. the routine
use of these methods alone to extract biomedical information is usually not recommended as the visualization is strongly dependent on the specific sample(s) examined. Without a comprehensive analysis of many samples and deeper understanding
of analysis at multiple levels of clustering, the effectiveness of these methods is
significantly reduced. A secondary source of knowledge is required to validate the
results of the classification, such as immunostaining or expert manual validation.
When used as a discovery tool, results can be further translated with supervised
methods to minimize errors and procure accurate results.
16.4 Application of Ft-IR in Histopathology, Especially
in Cancer
one of the major areas of application for IR imaging has been in diagnosing, staging
and grading cancer. A number of pioneering studies and reviews are available. For
the sake of brevity, we focus here on salient features of a few studies. For convenience, these largely focus on work from our own laboratories and are only meant to
be illustrative, not exhaustive or, necessarily, studies originating the ideas. We also
do not focus on straightforward applications, i.e. applications in which two sets of
tissues, one diseased and other normal, need to be separated. In our opinion, that is
a rather simple application that can be a byproduct of more comprehensive protocols and is of limited practical value in itself. hence, in our discussions of specific
studies, we focus on interesting aspects of translation, studies in which exciting new
information is available or where unexpected results were reported.
16.4.1 Breast Cancer
one of the most important applications of Ft-IR spectroscopy in analyzing breast
cancer tissues has been in trying to utilize Ft-IR automated histopathology for classifying the cell types present in breast cancer tissues, in order to diagnose, grade
and characterize the tumors. An example [50] of correlations between different cells
and stains and IR imaging data is shown in Fig. 16.10. IR absorption spectra have
been analyzed extensively for differentiating between the malignant and the benign
breast tissues [51–53].
S. Biswas et al.
Unsupervised Learning Common unsupervised learning methods include
K-means and hierarchical clustering analysis. In this type of classification, prior
knowledge of the classes within the data is not available and, indeed, the focus is
often on discovering the underlying structure. the data acquired is divided by the
algorithm used into clusters or groups. this form of classification is highly valuable
as a visualization and discovery tool in IR imaging. the results from unsupervised
methods may be used as is, especially when outliers are to be detected or a ready
reference (such as a corresponding stained tissue section) is available. the routine
use of these methods alone to extract biomedical information is usually not recommended as the visualization is strongly dependent on the specific sample(s) examined. Without a comprehensive analysis of many samples and deeper understanding
of analysis at multiple levels of clustering, the effectiveness of these methods is
significantly reduced. A secondary source of knowledge is required to validate the
results of the classification, such as immunostaining or expert manual validation.
When used as a discovery tool, results can be further translated with supervised
methods to minimize errors and procure accurate results.
16.4 Application of Ft-IR in Histopathology, Especially
in Cancer
one of the major areas of application for IR imaging has been in diagnosing, staging
and grading cancer. A number of pioneering studies and reviews are available. For
the sake of brevity, we focus here on salient features of a few studies. For convenience, these largely focus on work from our own laboratories and are only meant to
be illustrative, not exhaustive or, necessarily, studies originating the ideas. We also
do not focus on straightforward applications, i.e. applications in which two sets of
tissues, one diseased and other normal, need to be separated. In our opinion, that is
a rather simple application that can be a byproduct of more comprehensive protocols and is of limited practical value in itself. hence, in our discussions of specific
studies, we focus on interesting aspects of translation, studies in which exciting new
information is available or where unexpected results were reported.
16.4.1 Breast Cancer
one of the most important applications of Ft-IR spectroscopy in analyzing breast
cancer tissues has been in trying to utilize Ft-IR automated histopathology for classifying the cell types present in breast cancer tissues, in order to diagnose, grade
and characterize the tumors. An example [50] of correlations between different cells
and stains and IR imaging data is shown in Fig. 16.10. IR absorption spectra have
been analyzed extensively for differentiating between the malignant and the benign
breast tissues [51–53].