Page 40 - Vibrational Spectroscopic Imaging for Biomedical Applications

Towar d Automated Br east Histopathology 17

(a) 4 cm –1 (b) 1.0
Absorbance (offset for clarity) 16 cm –1 AUC at 8 metrics 0.9
–1
8 cm
–1
32 cm
–1
0.8
64 cm
0.7
0.6
Stroma
64
1000 1500 2000 2500 3000 3500 4000 0.5 4 8 16 32 Epithelium 128
–1
–1
Wavenumber (cm ) Spectral resolution (cm )
FIGURE 1.6 (a) Epithelium spectra are obtained by downsampling data
−1
acquired at 4 cm to lower spectral resolutions. (b) AUC analysis for stroma
and epithelium segmentation for each spectral resolution demonstrates a
decrease in classifi cation accuracy only at a very course spectral resolution.

resolution is downsampled to more course spectral resolutions using
a neighbor binning procedure. Average epithelium spectra (Fig. 1.6a)
demonstrate the effect of downsampling on spectral features. Impor-
tant spectral elements remain constant at 4, 8, and 16 cm resolu-
−1
tions, but peak locations and characteristic shapes begin to change
significantly at 32 and 64 cm . However, a significant drop in classi-
−1
fication accuracy does not occur until the spectral resolution decreases
−1
to 128 cm (Fig. 1.6b).
The robust classifier performance at downsampled spectral res-
olutions is permitted by the significant biochemical and spectral
differences between stroma and epithelium and the inherent nature
of the selected spectral metrics. As reflected in the spectra in Fig. 1.3,
numerous differences between these two tissue classes are visible
and indicate that there are significant biochemical differences
between these two types of tissue. Therefore, fine spectral resolu-
tion is not essential to distinguish stroma and epithelium. In addi-
tion, the peak height, area, and center of gravity metrics selected are
not extremely sensitive to small changes in spectral features. Spec-
tral absorbance values are generally measured accurately as long as
the full width at half maximum (FWHM) is not significantly less
than the spectral resolution. Therefore, many peaks are not affected
by moderate decreases in spectral resolution. Also, the center of
gravity metrics incorporated in the classifier depend on both peak
position and shape, and are therefore less significantly affected by
changes in peak location in downsampled spectra. The inherent bio-
chemical differences between epithelium and stroma and the types
of metrics selected for tissue segmentation allow the potential of
faster data acquisition at lower spectral resolutions without consid-
erable loss in classification potential.