Hyperspectral Imaging for the Differential Diagnosis of Squamous Cell Carcinoma and Keratoacanthoma of the Head and Neck.

Abstract

Accurate, label-free, non-destructive discrimination between head and neck squamous cell carcinoma (HNSCC) and keratoacanthoma (KA) remains challenging due to their overlapping morphology. We introduce a real-time, end-to-end hyperspectral imaging (HSI) workflow applied to 80 formalin-fixed, paraffin-embedded sections, each sampled with five 50 × 50-pixel ROIs and captured across 400-1000 nm to produce 128-band reflectance cubes. After reflectance calibration, Savitzky-Golay smoothing, and first-derivative preprocessing, a compact one-dimensional convolutional neural network achieved 87% accuracy, 93% sensitivity, 77% specificity, and AUC = 0.85 on a held-out test set. Spectral difference analysis revealed key biomarkers at the hemoglobin Q-band (630 nm) and OH overtone (917.5 nm), corresponding to vascular and extracellular matrix variations. This resource-efficient photonic platform enables rapid, automated "optical biopsy" without exogenous stains, offering scalable adjunctive diagnostics and a clear pathway toward intraoperative tissue classification.