Autoencoder-Based Detection of Nanoplastics in Biological Matrices via Infrared Hyperspectral Imaging

Le résumé fourni par la source

Abstract Mid-infrared hyperspectral imaging is an emerging tool for the qualitative and quantitative analysis of micro- and nanoplastics (MNPs) and for the characterization of biological tissues and complex matrices. Detecting MNPs in biological materials is of particular interest for assessing toxicological and ecological impacts; however, significant spectral overlap between polymer vibrational bands and those of proteins, lipids, and other biological components complicates identification at low MNP levels. In this work, an autoencoder-based anomaly detection approach is employed to learn the spectral characteristics of biological matrix signals from infrared spectra acquired with quantum cascade laser infrared (QCL-IR) microscopy. Residual-based anomaly mapping preserves chemically meaningful spectral features, enabling heatmap visualization of nanoscale plastic (NP) accumulation in two- and three-dimensional cell culture models. Fully connected (FC), convolutional neural network (CNN) and hybrid (CNN-FC) autoencoder architectures were evaluated, with FC and CNN-FC models providing a balance between accurate biological matrix reconstruction and preservation of NP spectral signatures. The method enabled detection of ∼50 nm plastic particles when present as localized accumulations corresponding to approximately 1% (m/m) of the dried biological material, equivalent to surface density on the order of 103 particles/μm2. These results demonstrate that residual anomaly detection can extend hyperspectral imaging to visualize nanoscale plastic accumulations in complex biological media at concentrations approaching the instrumental detection limit.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.