Page 39 - Vibrational Spectroscopic Imaging for Biomedical Applications

16 Cha pte r O n e

developed on the calibration TMA readily translates to a separate
spectral dataset. This indicates that the classifier does not overfit the
spectral data and has the potential to provide reproducible results in
a clinical setting.
Stroma and epithelium are easily visualized on classified images
(Fig. 1.5a) and appear to correspond accurately with H&E images
(Fig. 1.5b). The infiltrating tumors on malignant TMA cores are read-
ily recognized by the extensive green epithelium visible on the classified
image. The classified image provides the advantage of quick visualiza-
tion of tissue heterogeneity without the necessity of adding stains or
chemical dyes that irreversibly alter tissue properties. The qualitative
correspondence between H&E-stained images and classified spectral
images is confirmed by quantitative ROC analysis (Fig. 1.5c). This analy-
sis is performed after manual labeling of ~50,000 pixels in the validation
TMA spectral image as stroma or epithelium to serve as a gold standard
for classification evaluation. The small inset plot reveals that both stroma
and epithelium reach a maximum AUC of ~1 with the six metric classifi-
cation model obtained by validation on a separate dataset.
A comparison of the calibration and validation studies (Fig. 1.5d)
also indicates significant similarities in classification performance. In
both studies, the mean AUC curve reaches a maximum of over 0.99
with only six spectral metrics. A slight decrease in the AUC for each
curve past 60 metrics reveals that the data may be overfit by the addi-
tion of these metrics. This is even more noticeable in validation, which
is reasonable since the classification algorithm was not designed on
this dataset. A closer examination of the mean AUC curve for the first
six metrics in the plot inset indicates that the contribution of two met-
rics to classification performance is more significant in the calibration
than in the validation dataset. However, the other four metrics show
similar results in both calibration and validation.
Although this quantitative evaluation provides an excellent anal-
ysis of classification accuracy for stroma and epithelium as compared
with the gold standard tissue sections selected by a trained patholo-
gist, it does not provide any indication of tissue segmentation accu-
racy outside these selected regions of interest. This quantitative
analysis only evaluates supervised data, and does not provide a
numerical indication of the potential of this algorithm on unsuper-
vised spectral image classification. Image regions not included in this
quantitative evaluation include boundary pixels neglected in select-
ing regions of interest and infiltrating epithelial tissue sections inter-
mixed with malignant stroma in high-grade tumor samples, which
are difficult to manually select in tissue spectral images. Nonetheless,
a qualitative comparison of H&E and classified images indicates rea-
sonable classification of spectral image pixels that are not manually
mapped to a specific tissue class.
The dependence of classification accuracy on spectral resolution
is also considered. The validation data originally acquired at a 4-cm −1