Detection of contamination by aflatoxins on apricot kernels using NIR-hyperspectral imaging

Aflatoxins can be found as contaminants in a wide range of foods, such as nuts, cereals, dried fruits and milk. Due to their hepatotoxic and carcinogenic effects, the maximum allowed concentration of aflatoxins is nowadays regulated in many countries, with levels up to 50 μg/kg. In routine analysis, the main methods used to determine aflatoxins are based on high-performance liquid chromatography (HPLC) and enzyme-linked immunosorbent assay (ELISA) [1]. Despite the very high sensitivity of these methods, they are destructive, expensive, time consuming and not appropriate for real time control, e.g., online. Consequently, the development of fast, non destructive and economic methods for aflatoxins detection and monitoring in food industry is becoming more and more important. Our studies showed that manual sorting of dark or spotted apricot kernels removed 97.3-99.5% of total aflatoxins [2]. However, discolored seeds could be visually identified only after removing the skins from each seed by means of a time-consuming operation. For these reasons, in this work we investigated the possibility to use NIR–HSI for the fast and nondestructive automated identification of aflatoxin contaminated unpeeled apricot kernels. On the whole 9 hyperspectral images, each one containing 48 kernels, were acquired in the 900- 1700 nm range. After image acquisition, the kernels were peeled to identify the dark or spotted kernels and subjected to HPLC analysis for AFB1 quantification. Classification models were then calculated on a training set of NIR spectra extracted from a representative number of non-contaminated and dark seeds, selected on 5 images on the basis of HPLC analysis results as well as of the visual evaluation of the peeled kernels. The remaining 4 images were instead used as independent test set for model validation. Since dark seeds were found to have a higher concentration of AFB1 than spotted seeds, the latter ones were not included in the training set. Different iPLS-DA classification models, built using different signal preprocessing methods and different interval size values, were then evaluated in terms of classification efficiency in cross validation of the training set pixels, in order to select the optimal conditions. The results were reported under the form of predicted probability maps, and for each single kernel the contamination was estimated as the percentage of pixels assigned to the “contaminated” class by the iPLS-DA model.