Page 48 - Vibrational Spectroscopic Imaging for Biomedical Applications

Towar d Automated Br east Histopathology 25

the calibration TMA dataset and delays the sensitivity from reaching
a value of 1 on the ROC curve, resulting in a lower AUC value. How-
ever, the difference between the calibration and validation AUC val-
ues is demonstrated to be statistically insignificant due to the limited
number of cores on each TMA. In order to demonstrate a statistically
significant difference between the AUC values for these ROC curves,
each TMA would need to contain 10 times as many cancer and nor-
53
mal TMA cores. This indicates that although the number of TMA
cores included in this study is large enough to demonstrate the feasi-
bility of this algorithm for breast tumor discrimination, the limited
sample size for the study causes the quantitative ROC results to be
somewhat sensitive to outliers.
The preliminary study presented here provides evidence that
breast tissue spectral images segmented into stromal and epithelial
classes can be useful for tumor discrimination by the evaluation of
spatial information and epithelium content. ROC analysis indicates
that near-perfect discrimination of cancer and adjacent normal TMA
cores is possible by computing the fraction of simulated boxes rang-
ing in size from 6.25 × 6.25 μm to 75 × 75 μm that contain at least 50 percent
epithelium pixels. However, the results of this initial study provide
somewhat limited information about the application of this technique
to a large population of cancer patients, as the optimal cutoff for can-
cer segmentation remains unclear. Notwithstanding, this study indi-
cates that this method for cancer discrimination has potential for
automated tumor recognition. Further studies are required to pro-
vide a complete evaluation and optimization of this automated
method for tumor discrimination.

1.3 Conclusions
Recent technological developments in FT-IR spectroscopic imaging
have enabled the possibility for applications in histopathologic imag-
ing for cancer diagnosis and research. While many studies indicate
that FT-IR has the potential for clinical applications, at this time it has
not been adopted for automated histopathology in practice due to a
variety of factors. The primary reasons are a lack of robustly vali-
dated protocols in which data is acquired rapidly and classification is
efficient. The multivariate segmentation approach presented here
demonstrates promising results for reliable epithelium and stromal
recognition. The application of a breast TMA for data collection
allows for rapid acquisition and analysis of large datasets to select
robust classification metrics. This supervised classification approach
also has the advantage of providing insight into important biochemi-
cal properties of tissues by incorporating spectral metrics based on
tumor biological characteristics. This current research indicates that
FT-IR imaging may make a significant contribution to developing a
method for automated histopathology in a clinical setting.