Interpretable AI-driven hyperspectral imaging for rapid assessment of foodborne pathogen safety risks in fresh mutton

Foodborne pathogens in fresh mutton pose serious threats to human health, underscoring the need for rapid and accurate detection to ensure food safety. This study used short-wave near-infrared hyperspectral imaging (SWIR-HSI) combined with interpretable artificial intelligence to dynamically evaluate the importance of spectral wavelengths within a convolutional neural network (CNN) model. A novel interpretable AI framework, Target Specific SHAP Attention (TSSA), was proposed to quantify the contribution of each spectral band to pathogen classification through dynamically updated SHAP values, while integrating category-priority-based attention weights. This mechanism enabled precise focus on critical wavelengths and dynamically interpretable model decisions. Based on this approach, qualitative detection of foodborne pathogens contamination in mutton, including contamination identification and foodborne pathogen type recognition, was achieved. The TSSA-CNN model stably focused on critical spectral regions associated with the overtone and combination vibrations of CH, OH, and NH functional groups (1655.61 nm, 1435.68 nm, and 2309.26 nm) after 5400 training epochs. Ablation experiments confirmed that TSSA-CNN exhibited the best performance, with accuracies of 98.04% and 95.88% on validation and test sets, respectively. Moreover, compared with the baseline, TSSA significantly improved the recognition performance for high-priority categories, with relative accuracy increases of 16.00% for the "Unpolluted" category and 12.50% for the "Salmonella Typhimurium polluted" category. These findings demonstrate that integrating HSI with the TSSA-CNN model enables rapid and precise detection of foodborne pathogens in fresh mutton, providing a theoretical foundation for interpretable, target-specific, and dynamically adaptive pathogen monitoring technologies.