Separating spectral mixtures in hyperspectral image data using independent component analysis: Validation with oral cancer tissue sections

Recently, hyperspectral imaging (HSI) systems, which can provide 100 or more wavelengths of emission autofluorescence measures, have been used to delineate more complete spectral patterns associated with certain molecules relevant to cancerization. Such a spectral fingerprint may reliably correspond to a certain type of molecule and thus can be treated as a biomarker for the presence of that molecule. However, the outcomes of HSI systems can be a complex mixture of characteristic spectra of a variety of molecules as well as optical interferences due to reflection, scattering, and refraction. As a result, the mixed nature of raw HSI data might obscure the extraction of consistent spectral fingerprints. Here we present the extraction of the characteristic spectra associated with keratinized tissues from the HSI data of tissue sections from 30 oral cancer patients (31 tissue samples in total), excited at two different wavelength ranges (330 to 385 and 470 to 490 nm), using independent and principal component analysis (ICA and PCA) methods. The results showed that for both excitation wavelength ranges, ICA was able to resolve much more reliable spectral fingerprints associated with the keratinized tissues for all the oral cancer tissue sections with significantly higher mean correlation coefficients as compared to PCA (p < 0.001).