Chemometric Methods for Biomedical Raman Spectroscopy and Imaging - Emerging Raman Applications and Techniques in Biomedical and Pharmaceutical Fields - page 193

successes. Discriminant analysis has been used to analyze Raman spectra ex-
tensively in histopathology by Stone et al. [108, 109], Hench et al. [110], and
Puppels et al. [111, 112], in cervical [113], stomach [114] and colon tissue [115],
in human serum [116] and tears [117], foodstuffs [118, 119], in atherosclerosis
[120], for tissue engineering [121], animal models [122] of disease [123-125], cel-
lular activity [126] and animal products [127, 128]. A major common threat in
the applications is extensive pre-processing [129], a data reduction step (often
PCA) and frequent use of an exploratory unsupervised step [130], as illus-
trated in [131]. A number of authors have presented augmented algorithms
based on the LDA [132] and the application of these methods remains a focus
area, especially in spectroscopic analysis of disease.
8.3.5 Nearest-Neighbor Classifier
A simple classifier is based on a nearest-neighbor approach (Fig. 8.7) [133].
In this method, the closest sample from the training set to the sample
being classified is found using a scalar distance using the vector K . The object
being classified is assigned the same class as the class of the training object
so found. Nearest-neighbor methods have the advantage that they are easy
to implement. They provide remarkably accurate results if the features are
chosen carefully (and if they are weighted carefully in the computation of the
distance). Clearly, the similarity of vectors is based on the spectral feature
subset that is considered. In the absence of other tools or for rapid analysis
of data, nearest-neighbor methods are exceptionally useful. When representa-
tive spectra that are clearly distinct are encountered, these methods provide a
useful classification protocol. For a small number of classes and small training
sets, the methods work very well. A number of methods have been proposed to
calculate the distance between spectra or clusters of spectra, including mean,
median, and ensemble based to deal with noise and outliers.
There are also disadvantages of the nearest-neighbor methods. First, they
(like neural networks) do not simplify the distribution of objects in parameter
Fig. 8.7. Nearest-neighbor decisions involve the use of distance calculations between
the vectors of an unknown ( gray point at the center ) and moments of training data
( left )withlabels( black and light gray ) or more detailed calculations, for example,
average distance to all points ( right )