Deep Learning–Assisted identification and quantification of cell-associated microplastics using darkfield hyperspectral imaging

Microplastics (MPs) interaction with human cells poses potential health risks yet quantifying this process and its extensiveness remains challenging. In this study, we developed an automated strategy combining darkfield hyperspectral imaging (HSI) with a deep learning pipeline to detect and quantify cell-associated polystyrene (PS) microplastics at the single-cell level. This pipeline includes 3 main steps: the Mask R-CNN segmented individual cells, the least-squares support vector machine (LS-SVM) distinguished microplastic particles from cellular material, and the circular Hough transform (CHT) was used to count the particles detected within segmented cellular regions. This image analysis pipeline demonstrated high performance: Mask R-CNN detected cells with 95% precision, LS-SVM classified particle spectra with 99.7% accuracy, and the CHT detected particles with a precision of 99.6%. Together, these components enabled reliable quantification of particles within cells. Results showed a dose-dependent effect on the number of PS MPs in Caco-2 cells. At lowest concentrations (1 × 10 3 particles/mL), 21% of cells were detected with PS, while no impact on cell viability was observed. In contrast, at higher concentrations (1 × 10 8 and 1 × 10 9 particles/mL), 100% of cells were detected with PS and showed significant reductions in cell viability. Our findings demonstrated that integrating darkfield HSI with deep learning provides a robust quantitative assessment of MPs and cells interaction at single-cell resolution. This approach may be adaptable to other particle types and cell lines, subject to retraining and validation, offering a valuable tool for microplastic toxicology studies and complementing traditional high-throughput assays in evaluating the level of cell association and dose-response relationships.