Intelligent meat quality assessment using deep learning-integrated Vis/NIR spectroscopic devices: a comprehensive review

Visible and near-infrared (Vis/NIR) spectroscopy and hyperspectral imaging (HSI), combined with artificial intelligence, offer rapid and nondestructive approaches for assessing meat quality, freshness, composition, and authenticity. However, the reliability and industrial transferability of recently reported deep-learning (DL) models remain insufficiently evaluated. This review critically synthesizes different studies and compares their scope with directly relevant reviews, with emphasis on Vis/NIR device configurations, conventional chemometric and machine-learning methods, DL architectures, and their applications in meat-quality assessment. Particular attention is given to the fairness of DL–chemometric comparisons, biological sample independence, potential leakage arising from replicate spectra, pixels, or augmented observations, and the distinction between internal prediction sets and independently collected external validation. The reviewed evidence indicates that CNN-, LSTM-, attention-, and hybrid DL models can effectively exploit nonlinear and spectral–spatial information; however, their superiority over optimized conventional approaches is not universal. Several studies reported high internal predictive performance despite limited numbers of independent animals, carcasses, batches, or physical samples. Independently documented cross-site, cross-season, and cross-device validation was rare or could not be confirmed. Instrument variability, calibration transfer, environmental sensitivity, model interpretability, and prospective processing-line evaluation, therefore, remain major barriers to industrial adoption. Future research should prioritize group-aware data partitioning, transparent reporting of biological and derived sample numbers, representative multi-batch datasets, standardized DL–chemometric benchmarks, cross-device calibration transfer, stable biochemical interpretation of explainable-AI outputs, and prospective industrial validation. Overall, progress toward dependable AI-assisted meat spectroscopy will depend less on increasingly complex architectures than on rigorous validation, reproducibility, and demonstrated robustness under realistic operating conditions.