Researchers from Jiangsu University and Zhejiang University of Water Resources and Electric Power have developed a transfer learning approach that significantly enhances the accuracy and adaptability of NIR spectroscopy models for detecting mycotoxins in cereals.
A recent study explored a new method to improve mycotoxin detection in cereal grains. Led by Hui Jiang, a researcher at Jiangsu University, the study demonstrated the utility of near-infrared (NIR) spectroscopy and advanced deep learning strategies in improving the safety and efficacy of cereal grains. The study’s findings were published in the journal Food Chemistry (1). Jiang and his team, which was comprised of researchers from Jiangsu University and Zhejiang University of Water Resources and Electric Power, investigated how transfer learning, which is a machine learning (ML) technique, can dramatically improve the adaptability of spectral models used for food safety testing (1).
Cereals are an important ingredient in human nutrition. These foods, as long as they are not refined, are known to reduce the risk of some diseases, such as diabetes, coronary heart disease, and cancer (2). Cereal grains include wheat, barley, oats, corn, whole grains, and rice (2). However, cereals, when being cultivated and grown, are also susceptible to contamination by fungal toxins, or mycotoxins, such as zearalenone (ZEN) and aflatoxin B1 (AFB1). These toxins pose serious health risks to consumers, ranging from liver damage to cancer, and require rigorous monitoring to ensure food safety (1).
Currently, several detection methods have been developed using NIR spectroscopy for mycotoxin detection of cereal grains. However, although NIR spectroscopy showed promise, these methods often exhibited key limitations, including the poor adaptability of calibration models when applied to different instruments or grain types (1). This limitation significantly hinders broader implementation across different testing environments (1).
In this study, Jiang and his colleagues applied transfer learning strategies to chemometrics. The goal of their study was to improve the generalizability of deep learning models trained on spectral data, which they hoped would make these models more versatile and able to be used across different grain types (1).
Their study involved two different NIR spectrometers and data sets collected from wheat and peanut samples. The research team built their traditional quantitative models using Fourier transform near-infrared (FT-NIR) spectroscopy and standard NIR spectrometry. These models were then trained on spectral data to serve as source domain models, which were later adapted to new target domains using transfer learning techniques (1).
Evaluating the transfer learning models, the research team found that the second model performed better. What made the second approach successful, according to the team, is that it recalibrated the source domain model using partial data from the target domain, which significantly improved model accuracy (1).
For wheat samples contaminated with ZEN, the FT-NIR spectrometry model achieved a coefficient of determination (R²) of 0.9356 and a relative prediction deviation (RPD) of 3.9497 (1). For peanut samples containing AFB1, the R² reached 0.9419 with an RPD of 4.1551 using FT-NIR, while standard NIR yielded an R² of 0.9386 and an RPD of 4.0434 (1).
As part of the experimental procedure, nine transfer learning tasks were carried out to evaluate the methods under different scenarios. Importantly, the researchers also investigated how varying the size of the target domain training set impacted the final model performance (1). Their findings revealed that even small amounts of target data could substantially enhance transfer model accuracy, supporting the method’s practical value for real-world applications (1).
These results from this study show that transfer learning can be used to adapt spectral models to new scenarios, which goes a long way in resolving one of the key issues in cereal toxin detection. By recalibrating models with limited target domain data, the research team was able to significantly improve prediction accuracy across different grains and instruments (1).
As food safety laboratories worldwide strive for more standardized and efficient screening tools, transfer learning-enabled NIR spectroscopy models could become essential components in regulatory frameworks. The researchers propose that their method could lead to the development of centralized model libraries that are readily adaptable to different toxins, grain types, and detection equipment (1).
