Anatomical variability of in vivo Raman spectra of normal oral cavity and its effect on oral tissue classification

Abstract

We report the results of a clinical study to characterize the inter-anatomical variability of in vivo Raman spectra of normal oral cavity of healthy volunteers, and investigate its effect on the outcome of statistical discrimination of malignant and potentially malignant oral lesions from the healthy oral mucosa. An unsupervised cluster analysis using Fuzzy c-means clustering algorithm was conducted for quantifying the underlying structure of the normal oral tissue spectra. The algorithm was found to segment the normal oral tissue sites, based on similarity of spectral patterns, into four major anatomical clusters (AC): (1) outer lip, and lip vermillion border into AC-I with an accuracy of 80%; (2) buccal mucosa into AC-II with an accuracy of 72%; (3) hard palate into AC-III with an accuracy of 92%; (4) dorsal, lateral and ventral tongue and soft palate into AC-IV with an accuracy of 76%. A probabilistic multi-class diagnostic algorithm, developed for supervised classification, was used to classify the whole set of measured tissue Raman spectra into three categories: normal, potentially malignant and malignant. The results showed that the diagnostic algorithm, when applied on the pooled set of spectra from all the clusters, correctly discriminated normal, malignant and potentially malignant tissue sites with 86%, 88% and 86% accuracy respectively, which amounts to an overall accuracy of 87%. However, when the anatomy-matched data sets were considered, the overall classification accuracy was found to improve to 95% with the algorithm correctly discriminating the corresponding tissue sites with 94%, 99% and 91% accuracy respectively.