Wood2004

CREATED: 200906170724 LINK: url:~/Modules/Literature/Wood2004.pdf Established spectral patterns for different tissue types found in different stages of sqamous cervical epithelium.

Makes use of clustering, distance is computed using the correlation coefficient and clustering is done using neighbor-joining.

Reasonable noise does not affect clustering process, hence concludes that clustering is more stable than PCA, where increasing amount of noise is accumulated in the less relevant clusters.

Preprocessing

• removal of pixels with too high or low absorbance values or with poor SNR
• from the data set
• remaining spectra cut to 1800 to 800 cm^-1, smooted using SG algorithm and
• baseline corrected within this region
• vector normalized betwen 1800 to 800 cm^-1

Clustering performed on three spectral regions, 1800-800, 1740-1570, 1200-1000. Up to 10 clusters were selected corresponding to major anatomical features. For each cluster map the mean spectrum was extracted for comparison with other clusters.

Clustering using original intensities can be fine-tuned to reveal different tissue features. The use of spectral derivatives further increases the sensitivity of the cluster procedure and decreases dependence of the selected wave number range.

Derivative spectra emphasizes small differences in spectral band shape, rather than absolute intensities, thus small shifts in band positions and absence/presence of shoulders can be readily detected.