Page 30 - Vibrational Spectroscopic Imaging for Biomedical Applications

8 Cha pte r O n e

second, prior clinical knowledge is employed to guide the segmenta-
tion of tissue and the end result displays the data corresponding to
the chosen model without requiring any human intervention. An
example is a modified Bayesian classification 42,43 method in which
the universal set of pixel values is bounded by a prior clinical model.
The method necessarily requires the measurement of a large number
of known samples (priors) and, hence, was simply not possible in the
eras of mapping and early FT-IR imaging.

1.1.4 High-Throughput Sampling
While it is generally recognized that a large number of samples are
needed for calibration and validation of a prediction model, a conve-
nient method to image such large numbers was not available. The
development of tissue microarrays (TMAs) has provided a useful
44
solution and their application with rapid FT-IR imaging demon-
45
strated wide population robustness. TMAs can be obtained from
numerous tissue banks and incorporate small malignant and benign
tissue samples from many different patients on a single slide. This
tissue source promotes the development of large FT-IR imaging stud-
ies to achieve statistically significant histologic and pathologic clas-
sification results.

1.1.5 Modified Bayesian Classification and Automated
Tissue Histopathology
In this work, we follow an approach that combines the use of TMAs,
FT-IR spectroscopic imaging, and supervised automated histologic
46
segmentation. As outlined in Fig. 1.2, the experimental procedure
involves acquiring spectral images and examining tissue spectra to
select metrics for classification. The metrics are tissue spectral fea-
tures such as peak heights, ratios, areas, and centers of gravity. These
features capture the essential elements of the spectra, without regard
to histologic tissue type or disease state. Since the number of metrics
is considerably less than the number of spectral data points, this step
helps reduce the dimensionality of data and makes subsequent calcu-
lations fast. The next step is to determine the probability distribution
function (pdf) for each metric and quantitatively estimate the overlap
in pdfs. Pdfs are estimated from ground truth pixels that have been
marked manually by referring to a corresponding section that was
H&E stained and examined by a pathologist. The types of classes
marked by a pathologist are restricted to the task at hand. For exam-
ple, in this chapter, we report the two-class case in which epithelium
is first segmented from stroma. After histological epithelium recogni-
tion is established, the epithelium is further separated into benign
and malignant classes. Each cell type (class) is denoted by a false
color to provide visualization. The overlap in pdfs forms the region of
ambiguity in classification and provides a preliminary estimate of the