Page 44 - Vibrational Spectroscopic Imaging for Biomedical Applications

Towar d Automated Br east Histopathology 21

segmentation step, the method presented here for distinguishing can-
cer removes stromal image pixels prior to beginning cancer classifica-
tion. In addition, statistical analysis is only performed on epithelial
pixels manually selected by comparison with gold standard H&E
images. Therefore, the similarity in metrics selected in the optimized
classifiers for segmenting epithelium and stroma and segmenting
malignant and benign epithelium indicate potential epithelial to mes-
enchymal transition in tumor development. Conversely, less than
half of the metrics selected relate to DNA content and characteristics.
This indicates that focusing only on DNA biochemical features in
tumor spectra may not be the optimal method for distinguishing can-
cerous epithelial cells in intact human tissue.
These classification results for cancerous and adjacent normal
epithelium represent only an initial effort in cancer segmentation.
Further studies will be conducted to analyze the impact of spectral
noise on broadening malignant and benign probability distributions,
which in turn decreases segmentation accuracy estimates. Alterna-
tive segmentation methods such as genetic algorithms that do not
require prerequisite knowledge of metric population frequency dis-
tributions will be also considered for cancer classification, as varia-
tion between individual patients may make the application of a single
set of biochemical features to all tumors unfeasible. Finally, histologi-
cal discrimination of different types of stroma found in malignant
and benign tissue will be evaluated to analyze changes in stromal
features with malignant development.

1.2.6 Application for Patient Cancer Segmentation
Breast carcinomas are identified as a mass of epithelial cells. These
epithelial tumor masses can be discriminated from normal breast epi-
thelium by altered cellular morphology and tissue structure. We have
previously demonstrated automated prostate tumor discrimination
in spectral image datasets by evaluating altered cellular morphology
42
based on cell polarity. For breast tissue, we examine the potential for
tumor discrimination by evaluating the changes in tissue structure in
invasive tumors in which tumor cells move out of the breast ducts
and lobules and infiltrate surrounding tissue to create a large mass of
epithelium. Thus, malignant tissue contains more epithelium than
normal tissue, as previously noted in the discussion of the false-color
classified spectral images of TMA datasets (Fig. 1.5).
To quantify this change in epithelial tissue content, we have
developed a spatial analysis strategy termed “multiscale neighbor-
hood polling.” In this approach, boxes ranging in size from 1 × 1 pixel
(6.25 × 6.25 μm) to 12 × 12 pixels (75 × 75 μm) are evaluated by com-
puter simulation to compute the fraction of pixels classified as epithe-
lium for each box. This fraction is expected to be higher in cancer
TMA cores than adjacent normal TMA cores. To eliminate errors