Improving detection of Listeria monocytogenes in Queso fresco using generative adversarial network (GAN) augmented NIR spectral analysis.

Abstract

Rapid, non-destructive detection of Listeria monocytogenes (LM) in fresh cheeses is necessary for ensuring food safety, especially in high-risk products like Queso fresco (QF). In this study, NIR spectroscopy was combined with machine learning (Random Forest (RF) and Support vector machine (SVM)) and deep learning (1D CNN) approaches to classify LM population (0.0 log10 CFU/g, 1.0 log10 CFU/g, 2.0 log10 CFU/g, and 3.0 log10 CFU/g) and LM strain (ATCC® 43256™, ATCC® 43257™, and ATCC® BAA-3134™). To enhance the size of the dataset, spectral augmentation was performed using a conditional Wasserstein generative adversarial network with gradient penalty (cWGAN-GP), and model performance was evaluated using both real and augmented spectra. The results showed that GAN-based augmentation substantially improved strain-level classification performance compared with models that trained only on real spectra. Among the classification algorithms, SVM (95.7%) and 1D CNN (95.6%) achieved the highest classification accuracies on the augmented dataset, highlighting the benefit of synthetic spectral generation for improving robustness. Overall, this study emphasizes the potential of GAN-assisted NIR spectral augmentation along with machine learning (ML)/deep learning (DL) models as an on-site screening approach for LM detection and strain differentiation in fresh cheese.