Optimizing aflatoxin B1 detection in peanut kernels through deep modular combination optimization algorithm: A deep learning approach to quality evaluation of postharvest nuts

Abstract

Aflatoxin B₁ (AFB₁) is considered one of the most potent natural carcinogens. Quantitative detection of AFB₁ is essential for quality evaluation of postharvest nuts. In this study, a deep modular combination optimization (DMCO) algorithm was proposed to detect the content of AFB₁ in peanut kernels contaminated with Aspergillus flavus. The DMCO algorithm constituted a groundbreaking approach in the realm of deep learning for hyperspectral imaging analysis which meticulously selected and modularized existing deep learning models. It was characterized by the flexibility of combining these modules in serial configurations, parallel configurations or more complex configurations. This innovative architecture facilitated the capture of complex features, leading to improved predictive performance over single-module models. A performance-based selection mechanism was included in DMCO algorithm, which determined the most effective model architectures from a multitude of permutations. The optimal module combination reached a coefficient of determination for validation of 0.879, with root mean square error for validation and mean absolute error for validation recorded at 1.269 and 0.945, respectively. The DMCO algorithm successfully leverages deep learning to enhance the accuracy of AFB₁ detection in peanut kernels, showing its potential as a powerful tool to assess safety and quality for postharvest nuts.