Deep Learning Approaches for Oral Cancer Detection: A Comparative Study of Neural Networks and ResNet-18 CNN

Abstract

The prevalence of oral cancer, especially in South and Southeast Asia, coupled with late-stage presentation, is problematic globally as it profoundly increases the risk of death. Existing diagnostics, whilst effective, are invasive, lengthy, and often unavailable in areas with limited resources. This research aims to assess the viability of combining Raman spectroscopy and deep learning for oral cancer diagnostics and to evaluate if the new system would be efficient, accurate, non-invasive, and fast. To enhance data quality, Raman spectra obtained from oral tissues underwent baseline correction, normalization, denoising, and spectral alignment as preprocessing steps. An initial Neural Network (NN) and a ResNet-18 based Convolutional Neural Network (CNN) with transfer learning, were constructed and tested. It was proved that for every evaluation metric, the CNN model did better than the baseline NN, completing every evaluation metric with a better accuracy, precision, recall, F1 score, and ROC-AUC. All together, the results show deep learning, especially CNNs and transfer learning, can really help in oral cancer diagnostics and classification from complex spectral data, which is very useful for point-of-care diagnostics