Page 41 - Vibrational Spectroscopic Imaging for Biomedical Applications

18 Cha pte r O n e

1.2.5 Application for Cancer Pixel Segmentation
Upon successful differentiation of stroma and epithelium, the next
problem of automated segmentation of cancerous and normal epithe-
lium is addressed. All spectral image regions classified as stroma are
removed prior to commencing with cancer identification. Epithelial
pixels are then divided into cancer and normal classes based on iden-
tification of malignant tissue cores by a trained pathologist. These
manually selected epithelial pixels are used as a gold standard to
determine the frequency distribution of each spectral metric for can-
cerous and adjacent normal epithelium and to assess cancer identifi-
cation during statistical analysis. A malignant and benign epithelium
classifier is then calibrated using the procedure described in Fig. 1.2.
The pixel-based ROC analysis for calibration spectral image data
(Fig. 1.7a) reveals a maximum AUC of 0.79 with a nine-metric classi-
fier. This predicted accuracy for cancer and normal segmentation is
comparable with other pixel-based classification algorithms. 51
However, this classification accuracy is significantly lower in val-
idation spectral image data, as displayed in Fig. 1.7b. This is likely
due to several factors involved in determining segmentation accu-
racy. First, the spectral images were acquired with only two scans per
pixel to permit rapid data collection, resulting in a lower signal-to-
noise ratio (SNR). In addition, malignant tissues have vastly different
pathologic characteristics depending upon the organization of the
tumor, the type of tumor, and the grade of the tumor. This variation
in observed pathology would impact the biochemical features pres-
ent in the spectra. Therefore, attempting to group all malignant spec-
tra as one class may create difficulties in developing a reproducible
segmentation algorithm to separate all malignant tissue as a single
class. The metric probability distributions may not be uniform for dif-
ferent tumors and different patients, which would decrease predicted
accuracy confidence in the validation AUC curve. To further assess
classification potential, a core-based ROC analysis is performed to
analyze the sensitivity and specificity of the classification (Fig. 1.7c).
This resulting curve illustrates the trade-off between these two factors
in determining appropriate classification parameters. The threshold
for separating cancer and normal tissue could be altered based upon
a cost model for false positive and false negative results. The signifi-
cant decrease in AUC for cancer segmentation for validation data
(Fig. 1.7b) indicates that further studies to improve cancer detection
are required.
However, a comparison of calibration data H&E and classified
images indicates that using a threshold that assumes an equal cost for
false positives and false negative segmentation produces reasonable
cancer detection. As shown in Fig. 1.7d and e, many malignant tissue
cores contain a significant number of epithelial pixels classified as
cancerous while the normal tissue cores do not contain a significant
number of pixels classified as malignant. These images indicate that