Fluorescence Imaging and Machine Learning for Surface Contamination Detection: A Comparative Study with ATP Testing

Abstract Purpose This study evaluated UVC-induced fluorescence imaging as an alternative to ATP bioluminescence testing for detecting meat residues on food-contact surfaces. Objectives included assessing classification accuracy, correlation with ATP RLUs, key predictors, and model robustness. Methods and Results Fluorescence imaging at 275 nm and ATP assays produced 957 images, from which 23 intensity, texture, and edge features were extracted. A TabNetClassifier trained with cross-validation and SMOTETomek resampling achieved a Matthews correlation coefficient and Cohen’s Kappa of 0.77. GEE analysis showed no significant effect of resolution, gain, or exposure on sensitivity or specificity, with borderline impact for wood surfaces (p = 0.08). Texture features, especially Local Binary Patterns, were most predictive. Conclusion Fluorescence imaging with machine learning is a reliable alternative to ATP testing, offering consistent performance and minimal parameter sensitivity. Future work should refine calibration for specific surfaces and address practical deployment issues, including environmental variability, UVC safety, and integration with automated systems.