Infrared and Raman Spectroscopy and Imaging of Breast Cancer - Vibrational Spectroscopy for Tissue Analysis - page 135

It is also common to select a wavenumber range that is known or suspected
to pertain to the problem at hand, rather than using the entire spectrum. The
most commonly used range is the “fingerprint” region (~1000-1800 cm −1 );
the C-H stretching region (~2800-3000 cm −1 ) has also been used.
Classification approaches and algorithms
Most studies involve assigning a particular tissue type to every pixel in
the image via some kind of classification algorithm. K-means [60,74], fuzzy
C-means [59], and hierarchical clustering using Ward's method [59,68,71,72]
are popular unsupervised methods. Lasch [74] compared all three and
reported that hierarchical clustering was the most effective. In all of these
methods, the clusters are generated automatically, and then the resulting
colour-coded images are compared to the visible-light image of the same
sample. By searching for correspondences, clusters are assigned to tissue
types. These studies have focused on establishing the ability of infrared
microspectroscopy to discriminate between different kinds of tissue, but
have not looked at its use for actual diagnosis.
Studies that have attempted to use FTIR-MS imaging for diagnosis have
done so by using supervised classification methods, with a training set that
has been manually classified by a histological analysis. Several algorithms
have been used, including linear discriminant analysis (LDA) [61,62], arti-
ficial neural networks (ANNs) [71,72], soft independent modelling by class
analogy (SIMCA) [63], and a fast Bayesian classification scheme [58,64]. These
studies seek to develop predictive models that can be used to classify future
samples. In general, it seems to be relatively easy to distinguish between
tumour tissue and other types of tissue, but harder to distinguish between
different types of tumours [61]. However, there have been some examples of
successful tumour classification [63,72].
In some cases, PCA was used for data reduction prior to clustering [60,73].
In terms of data handling, this does not alleviate the main obstacle to cluster-
ing, which is the large number of objects (spectra). Clustering based on PCA
scores gives very similar results to using the raw spectra, unless very few
principle components (PCs; ~4 or fewer) are retained.
Some studies have used band-fitting approaches [57]. These are probably
quite labour-intensive and may be better suited for analysis of sample struc-
tural features within single images rather than for automated classification.
Fabian et al. [72] used a two-stage ANN method, in which the first neu-
ral network classifies a spectrum as being cancer or noncancer, and the sec-
ond network further classifies the cancer spectra. This approach is sensible
because the differences between cancer and noncancer are greater than the
differences among the different grades of cancer.
There are a number of factors that need to be considered for the set of data
reported here.