Location: Food and Feed Safety ResearchTitle: Near-infrared hyperspectral imaging for evaluation of aflatoxin contamination in corn kernels
|TAO, FEIFEI - Mississippi State University|
|YAO, HAIBO - Mississippi State University|
|HRUSKA, ZUZANA - Mississippi State University|
|KINCAID, RUSSELL - Mississippi State University|
|Rajasekaran, Kanniah - Rajah|
Submitted to: Biosystems Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/1/2022
Publication Date: 7/2/2022
Citation: Tao, F., Yao, H., Hruska, Z., Kincaid, R., Rajasekaran, K. 2022. Near-infrared hyperspectral imaging for evaluation of aflatoxin contamination in corn kernels. Biosystems Engineering. 221:181-194. https://doi.org/10.1016/j.biosystemseng.2022.07.002.
Interpretive Summary: Aflatoxin, produced by the fungus Aspergillus flavus is the most harmful mycotoxin and when ingested in high concentrations, can cause liver cancer in animals and human beings. Contamination at any level will result in loss or reduction in the value of the harvested product for the corn farmers. Our objective is to develop an innovative, rapid technology to detect aflatoxin contamination in corn kernels that is amenable for high throughput process and non-destructive sampling. We employed the near-infrared hyperspectral imaging (NIR-HSI) between 900 and 2500 nm wavelength to distinguish aflatoxin contamination on corn kernels in this study. Corn kernel samples inoculated with toxigenic or non-toxigenic fungi, Aspergillus flavus, were analyzed for aflatoxin content. Samples treated with non-toxin producing fungus were treated as negative reference class, Same kernel samples were also subjected to NIR-HSI in the laboratory. Statistical analyses of the results indicated that the overall accuracy of detection was 89.4 or 89.8% depending on the aflatoxin thresholds in the samples, 100 ppb or 20 ppb, respectively. Our results from this study indicate the advantage of using NIR-HSI (900 – 2500 nm) for identification of fungus-induced aflatoxin contamination on corn kernels and should be useful for the grain industry to rapidly evaluate aflatoxin contamination in the most important food and feed crop, corn in the United States and elsewhere in the globe to improve international trade and food safety.
Technical Abstract: Aflatoxins are among the most carcinogenic mycotoxins and are considered as the most harmful mycotoxins in U.S. agriculture. The innovative technologies that can offer the detection advantages of rapidness, non-destructiveness, and high-throughput are of urgent need. Therefore, the near-infrared hyperspectral imaging (NIR-HSI) between 900 and 2500 nm was employed to distinguish aflatoxin contamination on single corn kernels in this study. A total of 900 corn kernels were used with 3 treatments, namely, 300 kernels inoculated with the AF13 fungus (aflatoxigenic), 300 kernels inoculated with the AF36 fungus (non-aflatoxigenic), and 300 kernels inoculated with sterile distilled water (control). One hundred kernels from each treatment of 300 kernels were subjected to incubation at 30 °C for 3, 5 and 8 days, dried and wiped to remove exterior signs of mold, and imaged. The kernels were imaged on both endosperm and germ sides, and the reference a'atoxin concentration in each kernel was determined by the VICAM AflaTest method. The mean absorbance spectra extracted from single kernels were preprocessed by standard normal variate (SNV), first derivative (FD) and second derivative (SD) transformations. The original and preprocessed endosperm-, germ-side and mean of both sides were used as inputs for establishment of partial least-squares discriminant analysis (PLS-DA) models. The full spectral PLS-DA results show that the models established upon the SNV-preprocessed germ-side absorbance, generally performed better in discriminating aflatoxin contamination. The modeling results with different training sample size ratios in terms of negative: positive class, indicate that the 2:1 ratio achieved more balanced prediction performance for both negative and positive classes. With the classification threshold of 20 ppb, the type II- competitive adaptive reweighted sampling (CARS)-PLSDA models which were established upon 17 commonly selected variables under the condition of “frequency = 70” achieved comparable prediction capability to the corresponding full spectral and type I-CARS-PLSDA models. Based upon the 20-ppb threshold, the mean prediction accuracies of the negative class, positive class and overall accuracy of the 100 random runs of type II-CARS-PLSDA models, were 90.4%, 82.2% and 89.8%, along with the SDs of 1.4%, 5.5% and 1.3%, respectively. When 100 ppb was applied as the classification threshold, 21 commonly selected spectral variables under the condition of “frequency = 60” were needed for the type II-CARS-PLSDA models to achieve similar prediction performance to the full spectral and type I-CARS-PLSDA models. Based upon the 100-ppb threshold, the mean prediction accuracies of the negative class, positive class and overall accuracy of the 100 random runs of type II-CARS-PLSDA models, were 89.8%, 83.0% and 89.4%, along with the SDs of 1.5%, 6.7% and 1.4%, respectively.