Page 34 - Vibrational Spectroscopic Imaging for Biomedical Applications

Towar d Automated Br east Histopathology 11

paraffin removal, fresh hexane is added every 3 to 4 hours. Paraffin
elimination is checked at 24 hours to monitor the disappearance of
−1
the 1462 cm peak on several tissue cores.
A Perkin-Elmer Spotlight 300 spectrometer is used for collection
−1
of spectral images at a 6.25 μm pixel size and a 4 cm nominal spec-
tral resolution with 2 scans per pixel. An IR background is collected
at 120 scans per pixel at a location on the array substrate with no
sample present. Tissue spectral images are output as the ratio of the
raw data to the background spectra. Spectral images are then com-
piled, analyzed, and classified using Environment for Visualizing
Images (ENVI) imaging software with programs written in the inter-
active data language (IDL) compiler to perform the classification
analysis described in Fig. 1.2.

1.2.1 Models for Spectral Recognition and
Analysis of Class Data
The first model developed for breast tissue classification involves the
segmentation of stroma and epithelium. This step is necessary to
determine the important spectral features for breast tissue, as these
47
are two of the most prominent tissue classes in the breast. Distin-
guishing epithelium from stroma is particularly important, as over 99
48
percent of malignant breast tumors arise in epithelial tissue. There-
fore reliable epithelial segmentation is a prerequisite for tumor recog-
nition. Stromal identification is also important as many recent cancer
studies have highlighted the importance of the stromal microenvi-
49
ronment in epithelial tumor development. As stroma and epithe-
lium display significantly different biochemical properties, they
should be segmented in spectral images with a high degree of classi-
fication accuracy and confidence.
To develop a classification model, spectral image regions are
identified by comparing FT-IR and H&E images to select spectral
image pixels that clearly correspond to stroma or epithelium. Approx-
imately 200,000 pixel spectra are selected for calibration to eliminate
errors due to variation between individual patients and inherent
spectral noise. Selection of a large number of spectra for calibration is
necessary to ensure classification accuracy in validation studies.
Average spectra for stroma and epithelium are then computed from
these selected spectral image pixels and are displayed in Fig. 1.3.
Important spectral features are selected by examination of tissue class
average spectra to reduce data dimensionality prior to classification.
Prominent stroma and epithelium spectral features are compared
with breast tissue spectral features previously identified by other
groups 33,39,40 to assess the biological relevance of each metric. Stroma
and epithelium spectra are then distinguished by considering spec-
tral features associated with unique biochemical tissue properties.
For example, epithelial tissue is observed to have a higher relative