Application of machine learning for feature-subset selection in spectroscopic data is valuable for label free spectral-marker identification in disease diagnosis. Reduction of both computation and healthcare cost is main aim of such endeavor. In this study, Fourier Transform Infrared spectroscopy (FTIR ) has been used for computationally optimal delineation of oral squamous cell carcinoma (OSCC) from normal oral mucosa (NOM), oral leukoplakia (OLK) and oral sub-mucous fibrosis (OSF) as two-class classification problem using features obtained through forward feature selection, a greedy wrapper based computational stepwise feature reduction technique without transforming the existing features to obtain uncorrelated and discriminative features. In each case six spectral markers were obtained using open-source IRootLab toolbox for vibrational spectroscopy in MATLAB R2015a. 57 tissue biopsy samples (7 NOM, 11 OSF, 16 OLK and 23 OSCC) were collected. Four micrometer thick unstained histologically diagnosed tissue sections were used for KBr pellet preparation. FTIR spectra were obtained for the range of 400-4000 cm-1 at a resolution of 4 cm-1 with 32 scans, using an 8-mm aperture diameter and deuterated triglycine sulfate detector. One spectrum per sample for three sections of each subject in transmission mode was acquired. Result implied that the lesions could be completely segregated, when linear discriminant analysis (LDA) scores plot with confidence ellipse representing 95% confidence interval of preprocessed region between 900–1800 cm-1 were plotted after principal component analysis for computationally feature reduction of dimensionality before LDA, suggesting significant variation of chemical composition in oral lesions. 1032, 956, 1707, 1639, 1606, and 1565 cm-1 were found to be significant in NOM-OLK, while 1782, 1713, 1665, 1545, 1409, and 1161 cm-1 in OLK-OSCC, 1687, 1619,1531,1481,1384, and 1322 cm-1 in OSF-OSCC and 1670,1306,1757,1723,1611,1554 cm-1 in OLK-OSF were found to be the selected spectral-markers from generated feature histogram. It can be concluded that such alteration in chemical signatures could be mainly due to protein phosphorylation, and suggested computational technique is capable of theoretically relevant peak picking for optimal diagnosis oral pre-cancers, with improved generalization capabilities and reduced run-time.

Fourier transform infrared spectroscopic spectral feature subset selection for optimal diagnosis of oral lesion

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.