Deep convolutional neural networks for tongue squamous cell carcinoma classification using Raman spectroscopy

Abstract

With deep convolutional neural networks and fiber optic Raman spectroscopy, this study presents a novel classification method that discriminates tongue squamous cell carcinoma (TSCC) from non-tumorous tissue. To achieve this purpose, 24 tissues spectral data were first collected from 12 patients who had undergone a surgical resection due to the tongue squamous cell carcinomas. Then 6 blocks with each block including 1 convolutional layer and 1 max-pooling layer are used to extract the nonlinear feature representations from Raman spectra. The derived features form a representative vector, which is fed into a fully-connected network for performing classification task. Experimental results demonstrated that the proposed method achieved high sensitivity (99.31%) and specificity (94.44%). To show the superiority for the ConvNets classifier, comparison results with the state-of-the-art methods show it had a competitive classification accuracy. Moreover, these promising results may pave the way to apply the deep ConvNets model in the fiber optic Raman instrument for intra-operative evaluation of TSCC resection margins and improve patient survival.

Keywords: Convolutional neural networks (ConvNets); Deep learning; Fiber optic raman; Raman Spectroscopy; Spectroscopy; Tongue squamous cell carcinoma.