Hyperspectral image analysis of oral squamous cell carcinoma using machine learning techniques

• Ano de defesa: 2025
• Autor(a) principal: Peres, Daniella Lúmara Pereira Mendes de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: aprendizado de máquina; câncer oral; espectroscopia FTIR; espectroscopia vibracional; FTIR spectroscopy; hyperspectral imaging; imagem hiperespectral; machine learning; oral cancer; vibrational spectroscopy
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/85/85134/tde-11122025-123821/

Oral squamous cell carcinoma (OSCC) remains one of the most aggressive malignancies of the head and neck region, with prognosis heavily dependent on early detection. Hyperspectral imaging (HSI) combined with Fourier Transform Infrared (FTIR) spectroscopy is capable of capturing detailed biochemical information from tissue samples. In this study, we investigated the performance of four machine learning (ML) models - Linear Discriminant Analysis (LDA), Partial Least Squares Discriminant Analysis (PLS-DA), Random Forest (RF), and Feed Forward Neural Networks (FNNs) - for the classification of FTIR hyperspectral images of OSCC and healthy oral tissue. Human tissue microarray samples, comprising 48 OSCC and 48 control specimens, were preprocessed using spectral trimming, smoothing, Extended Multiplicative Signal Correction (EMSC), and Standard Normal Variate (SNV) normalization. Spectra were unfolded for pixel-level analysis, and classification performance was evaluated through 10-fold cross-validation (CV) using metrics such as accuracy, F1-score, and the area under the ROC curve (AUC). LDA achieved robust results at both pixel and image levels, with an AUC of 0.9465 and 91.7% image-level accuracy. PLS-DA demonstrated strong pixel-level classification (AUC = 0.8686) but showed decreased performance at the image level. Random Forest outperformed the other models in pixel-level analysis (AUC = 0.9864) and maintained satisfactory image-level performance. FNNs achieved balanced accuracy (80%) and high-lighted spectral regions related to protein secondary structures as key discriminators. These findings confirm the potential of FTIR-HSI coupled with ML as a powerful tool for the early diagnosis of OSCC, with LDA and RF models offering particularly favorable performance in both interpretability and predictive capability.