Tongue squamous cell carcinoma discrimination with Raman spectroscopy and convolutional neural networks

Abstract

Discriminating the border of tongue squamous cell carcinoma (TSCC) is very critical for surgical treatment. Based on fiber optic Raman spectroscopy and deep learning technique, this study proposed a framework with convolutional neural networks (CNN) to discriminate TSCC from non-tumorous tissue. First, Raman spectra of 24 samples of TSCC and adjacent tissues from 12 patients were collected by fiber optic Raman spectroscopy system. Through the analysis, the significant differences between TSCC and non-tumorous tissue were occured in the range of 700–1800 cm−1. Then a CNN model was used to extract the nonlinear feature representations from Raman spectra. Finally, extracted features are fed into a fully-connected layer for TSCC classification. The results demonstrated that the CNN model obtained the sensitivity and specificity of 99.07% and 95.37%, respectively. Moreover, comparison results with existing methods showed our method achieved the highest accuracy of TSCC classification. Therefore, Raman spectroscopy combined with the CNN model has a great potential to provide a useful technique for the intraoperative evaluation of the margin of resection of oral tongue squamous cell carcinoma.