Espectroscopia ATR-FTIR associada a algoritmos de inteligência artificial para discriminação de nódulos benignos e malignos em tecido tireoidiano

Abstract

Thyroid nodules are commonly detected in clinical practice, and while most are benign, ruling out thyroid cancer remains a primary concern. The diagnosis can be challenging, as cytology from Fine Needle Aspiration often yields inconclusive results, potentially leading to a diagnostic thyroidectomy. The present study assesses the potential of Attenuated Total Reflection-Fourier Transform Infrared spectroscopy combined with machine learning to differentiate benign from malignant thyroid nodules in fresh frozen tissue samples. Thyroid nodular tissue samples were collected from patients undergoing thyroidectomy due to suspected malignancy. Infrared spectra were then acquired from these tissue samples, with preprocessing and analysis focusing on the lipidic (3050–2800 cm⁻¹) and biofingerprint (1800–800 cm⁻¹) regions. These infrared spectra were subjected to a series of machine learning algorithms, including Support Vector Machine, Random Forest, AdaBoost, Artificial Neural Network, Gradient Boosting and Naïve Bayes. The Gradient Boosting algorithm achieved the highest performance of discrimination, achieving an accuracy of 96%, a sensitivity of 100%, and a specificity of 92%, indicating capability to effectively differentiate between both types of nodules. SHAP analysis revealed key wavenumbers (985, 1561, and 2904 cm⁻¹) as crucial for spectral differentiation, correlating with biomolecular alterations in neoplastic tissue. These findings highlight the potential of Attenuated Total Reflection-Fourier Transform Infrared spectroscopy as a rapid, reagent-free and sustainable classification method, providing a proof of concept for differentiating benign from malignant thyroid nodules in fresh-frozen tissue samples.