Deep Learning Classification of Oral Cancer Using µFTIR Amide Band Images

Abstract

Oral squamous cell carcinoma (OSCC) is the most common malignancy of the oral cavity and is frequently diagnosed at advanced stages. Protein remodeling is a hallmark of malignant transformation, making the Amide I, II, and III spectral regions promising biomarkers for OSCC detection. Fourier-transform infrared (FTIR) hyperspectral imaging combined with convolutional neural networks (CNNs) enables automated tissue classification from biochemical images.
FTIR hyperspectral images were acquired from tissue microarray (TMA) samples containing OSCC and control tissues. Spectra were preprocessed using signal-quality filtering, Standard Normal Variate (SNV) normalization, and Savitzky–Golay second derivatives. Only the Amide I, II, and III regions were retained, and each selected wavenumber was converted into a two-dimensional intensity image. Specimen-level partitioning prevented data leakage, and data augmentation was applied only to the training set. A CNN was trained using the Adam optimizer and binary cross-entropy loss. Performance was evaluated at image and specimen levels, with specimen classification based on majority voting.
The proposed method achieved accuracies of 83.3% and 82.6% at image and specimen levels, with ROC-AUC values of 0.924 and 0.920, respectively. Correct classification was obtained for 13 of 16 OSCC specimens and 6 of 7 controls, demonstrating that amide-band images provide sufficient biochemical information for accurate OSCC classification while reducing spectral dimensionality and computational cost.

Funding: CNPq (INCT-INTERAS 406761/2022-1, PQ 314517/2021-9); Sisfoton (440228/2021-2); CAPES (88887.176297/2025-00; Finance Code 001); FAPESP LOOPS (2022/0355-9).

Speaker

Daniella Lúmara Pereira Mendes de Oliveira Peres
Nuclear and Energy Research Institute & University of São Paulo
Brazil

Discussion

Ask question