Page 29 - Vibrational Spectroscopic Imaging for Biomedical Applications

Towar d Automated Br east Histopathology 7

connective tissue. FT-IR imaging with array detectors also allows for
spatial resolution near the cellular level, which provides opportuni-
ties for detailed tissue analysis. This simple example demonstrates
the applicability of chemical imaging in distinguishing prominent
tissue features without the use of chemical dyes or contrast agents,
yet in a manner that is appreciable by practitioners who may not be
experts in spectral analysis.
FT-IR imaging has been demonstrated to be a useful tool in the
18
analyses of many tissue types. For breast tissue and cancer, a num-
ber of studies have provided evidence of feasibility. One of the first
successful efforts involved a cohort of 77 breast tumor samples and
incorporated linear discriminant analysis with cross validation to
20
classify tumors by grade and steroid receptor status. While the clas-
sification results were fairly accurate (87 percent for tumor grade and
93 percent for steroid receptor status), the study lacked independent
validation data. Further translational activities were not reported to
establish classification in a clinical setting. Several subsequent trials
conducted by another group involved a collection of several thou-
sand spectra from approximately 25 breast cancer patients with fibro-
adenoma, DCIS, or invasive ductal carcinoma. A supervised artificial
neural network (ANN) analysis was used to develop an automated
40
classifier. In a separate study using a subset of the same data, clus-
ter analysis was performed on 96 spectra in the fingerprint spectral
34
range to separate fibroadenoma and DCIS. Although these results
show good classification accuracy, a much more extensive study is
needed to evaluate the diagnostic potential of this algorithm and
ensure that calibration data is not overfit by the ANN. Further, it is
difficult to interpret the results of a complex, nonlinear classifier.
Other approaches to classify breast tissue involved the novel use of
slides and staining, as practiced in clinical settings to ensure compat-
ibility to current practice. 34,41 The results were promising in small
cohorts but larger efforts are needed to ensure that the promising
early results can provide a consistent and practical protocol for clini-
cal translation.
While many of these studies applying FT-IR spectroscopy for dis-
ease diagnosis have produced interesting results, they have not
widely attracted attention from clinicians due to their preliminary
nature. Specifically, the small numbers of patients included in these
studies and the difficulties in achieving effective large-scale valida-
tion results due to overfitting small sets of spectral data remain con-
cerns. Various approaches to the microscopic analysis of tissue
structure have been used but can be divided roughly into two major
categories. In the first, spectroscopy is used to guide the visualization
of tissue. For example, displaying pixels with spectral similarity with
the same color codes allows a human to recognize gross structure. An
example of this approach is a hierarchical clustering analysis that
allows a user to choose the number of clusters to be displayed. In the