By integrating the GEPARD software with a custom-trained machine learning model, Imptox researchers from HES-SO have made microplastic detection faster and more accurate.

The plastic pollution problem is more than meets the eye - literally. While we can see plastic debris littering our beaches and landscapes, some plastic particles are too small to be visible. Microplastics, tiny particles smaller than five millimetres, and even smaller nanoplastics have infiltrated our food, water, and the air we breathe. Scientists, including those at Imptox, are increasingly studying the effects of these tiny plastics on human health, especially as they are now being detected inside the human body. But how do you track what you can’t see?

Making MNP Analysis Faster with Machine Learning

That’s where machine learning comes in. In this study, Imptox researchers Thibault Schowing, Carlos Peña-Reyes, and Xavier Brochet from the Haute école spécialisée de Suisse occidentale (HES-SO) sought to make the detection and analysis of micro- and nanoplastics (MNPs) faster and more efficient. By improving an existing machine learning tool, they have adapted it to better identify MNPs in environmental samples. The goal was to reduce the time and manual effort typically required to detect tiny plastic particles, making the process more efficient and adaptable across various sample types.

Challenges with Existing Software

The researchers worked with a Nicolet™ iN10 Infrared Microscope, which allows for the Fourier Transform Infrared (FTIR) analysis needed to quantify and characterize micro- and nanoplastics (MNPs) in environmental samples. The microscope, paired with the OMNIC™ Picta™ software, automates the collection of composed images and provides wizards to assist in particle detection and identification. While the existing software is effective for many standard detection tasks, it involves some manual steps that can slow down the process when analysis needs to be done for hundreds of samples. For example, separate detection processes are needed for darker and lighter particles, which adds time and bias to the analysis. Additionally, in samples with a high concentration of tightly clustered particles, it can be challenging to achieve optimal results.

Switching to Open-Source Software GEPARD

To address these functional gaps, the research team opted to use the open-source software GEPARD, developed by the Leibniz Institute of Polymer Research Dresden. GEPARD provides additional control and flexibility for handling a wide variety of sample types, as well as a platform that supports the integration of machine learning, enabling faster and more automated detection of microplastics. GEPARD processes images efficiently in a single step across diverse particle types, providing enhanced options for sample preparation and data acquisition.

Training a Custom Neural Network for Better Detection

The researchers took GEPARD a step further by integrating a custom-trained Artificial Neural Network (ANN) - a type of AI designed to recognize patterns in data. While GEPARD already had an ANN designed for Raman images, these models weren’t suitable for detecting microplastics using FTIR images. To address this gap, the HES-SO team used images received from the University of Belgrade's FTIR group to create a usable dataset. This dataset was then used to retrain the neural network to handle FTIR data effectively, enabling more accurate detection of microplastics, even in complex or heavily saturated samples.

The retrained neural network now works alongside GEPARD to automate the detection process. This innovation allows the tool to identify the shape, size, and composition of microplastics more quickly and reliably than before.

Better Performance and More Accurate Results

Once the custom-trained neural network was integrated into GEPARD, the team compared its performance with the existing OMNIC™ Picta™ software. The results showed a significant improvement: GEPARD detected about 15% more particles than OMNIC™ Picta™ under similar conditions. This means the new method not only works faster but also provides more accurate and thorough results, making it easier to detect smaller and more complex particles that might have been missed with the older software.

All Data and Methods Available in Open-Access Format

In addition to the technical improvements, the researchers made all the data and information needed to replicate their process available in a fully open-access format on Zenodo. They detailed the setup of their software environments and the steps they took to train the neural network, using popular tools like Detectron2 and PyTorch. This ensures transparency and makes the process easily accessible to other researchers, helping to spread these advancements to the wider scientific community.