Page 275 - Vibrational Spectroscopic Imaging for Biomedical Applications

Raman Micr oscopy for Biomedical Applications 251

the database. Here, results obtained from SVMs will be presented.
During the SVM training functions are found which divide the
classes (spectra of bacteria from different strains) with a maximum
margin. These decision boundaries between the classes are
described by so-called support vectors. A robust training of the
data set can be achieved with (the time-consuming) leave-one-out
testing. That means, all spectra, except one spectrum, are used for
calculating the model and the left-out spectrum is used for classifi-
cation. Then, the same procedure is repeated for another data point
to be left out, until every object was removed once from the train-
ing set. For the subsequent classification of new data not the whole
training dataset is needed anymore, but only the support vectors
are used. 54,55
Bacterial strains which could be found in a clean room environ-
ment were chosen to create a database using support vector machines
with a nonlinear gaussian radial basis function (rbf) kernel. These
spectra include the spectra shown in Fig. 8.10a plus additional bacte-
rial strains within same species. To ensure a stable classification, bac-
teria grown under varying environmental conditions were included
56
in the training set. The results of the SVM classification of the
Raman spectra from the single bacterial cells excited at 532 nm are
shown in Table 8.1. Altogether the dataset consists of 2545 spectra from
20 strains of nine species. The bacteria were identified with recogni-
tion rates of 96 percent on strain and with 97 percent on species
level. 53,56,57
Once the database is created new single bacterial cells can be
investigated within a minimum of time. Table 8.2 shows the clas-
sification results for an independent set of different single bacteria
from 16 different strains. The spectra were recorded anonymously
and classified with the previously created SVM. One hundred and
twenty-five out of 130 bacteria were correctly identified. Out of the
five incorrectly classified spectra four were assigned to the wrong
strain within the right genus, and only one was misclassified on the
species level. This proves the high potential of automated Raman-
based assignment. Single Raman spectra acquired with 60 seconds
exposure time enable to assign 96 percent of the bacteria on the
58
strain level and 99 percent on the species level. A prototype of a
fully automated device for the fast, reliable, and nondestructive
online-identification of microorganisms from clean room environ-
59
ment has been manufactured. The approach utilizes fluorescence
labeling of single bacterial cell to differentiate live/dead bacteria or
60
biotic/abiotic particles prior the Raman identification step. First
instruments which could speed up the identification of microbial
pathogens are already on the market (SpectraCell RA™ Bacterial
Strain Typing Analyzer) or under development (rap.ID Particle Sys-
tems for the analysis of microparticles from fluids or air). These
classification methods can also be extended to eukaryotic cells. 55