AI and Light-Based Imaging Take Aim at Microplastics in the Food Supply

Fish swimming in water polluted by plastics

Microplastics have become a near-ubiquitous presence in the natural world, turning up in oceans, soil, drinking water and increasingly in the food supply. Detecting them in seafood, however, has long required a laborious and destructive process: chemically digesting or otherwise breaking down the sample to isolate the plastic particles before they can be identified.

Working together with colleagues at Virginia Tech, Lifeng Zhou, PhD, an assistant professor of electrical and computer engineering at Drexel University, has developed a way to detect microplastics directly on fish surfaces without destroying them, by combining light-based imaging with artificial intelligence. The study appears in the Journal of Chemometrics.

The technique at the center of the research is hyperspectral imaging, which captures a continuous spectrum of light intensity across hundreds of wavelengths for every pixel in an image, integrating spatial and spectral information in a way that enables precise material identification. The approach has gained traction in food safety research because it requires minimal sample preparation and is well-suited for field applications. The challenge in detecting microplastics on food, however, is that the spectral profiles of common plastic polymers like polyethylene overlap significantly with those of organic food components such as proteins, water and lipids, limiting the effectiveness of conventional analysis.

To overcome that challenge, the team trained neural network models to classify each pixel in a hyperspectral image as either microplastic or fish tissue. They contaminated tilapia fillet samples with small polyethylene microspheres, 300 and 600 micrometers in diameter, and scanned them using a near-infrared hyperspectral imaging system. The best-performing approach, a one-dimensional convolutional neural network (1D-CNN) applied directly to the full spectral data, achieved detection scores of 0.963 for the larger particles and 0.950 for the smaller ones.

"The convolutional architecture is particularly well-suited to this problem because it learns from localized regions of the spectral profile, capturing correlations between adjacent wavelengths rather than treating each one independently," Zhou said. "That sensitivity to local patterns is what allows it to distinguish microplastics from fish tissue even when the two look nearly identical at a glance."

That result held up only when the neural network was given access to the complete spectral data. The team tested whether applying principal component analysis, a common technique for simplifying large datasets, would improve performance. It did not. Compressing the spectral data before training consistently lowered detection accuracy, suggesting the model relies on fine-grained spectral details that dimensionality reduction tends to discard.

The findings point toward a practical path for food safety screening that sidesteps the need for sample destruction. Current quality-control methods depend on isolating microplastics through chemical digestion and examining them under microscopes or spectrometers, a slow process requiring specialized equipment. An imaging-based approach could eventually be adapted for use at processing plants and fish markets without extensive sample preparation.

"Microplastic contamination in seafood is a problem regulators are still working out how to address, and part of what makes that hard is the absence of fast, standardized detection methods," Zhou said. "Establishing those methods is really the next frontier, and we see this kind of imaging-based approach as a foundation for that work."

The researchers note that the current study was conducted under controlled laboratory conditions and focused on a single plastic type. Future work will aim to expand the training dataset to include more varieties of seafood, other plastic polymers commonly found in marine environments, and samples with varying moisture levels, which can complicate spectral readings.

Read the full study at https://doi.org/10.1002/cem.70088


In This Article