Oral cancer diagnosis via machine learning and micro-FTIR hyperspectral imaging

Abstract

Oral squamous cell carcinoma (OSCC) remains a major global health challenge, and its early detection is essential for improving patient prognosis. Fourier-transform infrared (FTIR) hyperspectral imaging offers a powerful, label-free approach for probing biochemical alterations in biological tissues, providing rich spectral information that can support computational diagnosis. In this study, we evaluated the performance of a Random Forest classifier applied to FTIR hyperspectral images of OSCC and control oral tissues. After rigorous preprocessing and restriction of the analysis to the 1500-1750 cm−1 region—the portion of the fingerprint band with highest discriminative variability—the model achieved strong pixel-level performance, with high precision, recall, and an area under the ROC curve of 0.986. Image-level classification, a clinically relevant metric, yielded an accuracy of 0.86 across 96 samples, demonstrating the model’s reliability in assigning whole-sample labels. Feature importance analysis identified key vibrational modes associated with malignant transformation, notably the Amide I and Amide II bands, as well as lipid-related C=O stretching near 1735 cm−1. These findings indicate that Random Forest, combined with FTIR hyperspectral imaging, can robustly capture biologically meaningful spectral signatures and represents a promising complementary tool for OSCC screening and diagnostic support.