Improving Quantitative Reliability of Single-Particle ICP−MS in Ion-Rich Backgrounds Using Deep Learning-Assisted Signal Reconstruction

Abstract Single-particle inductively coupled plasma mass spectrometry (spICP-MS) has been widely employed for quantifying nanoparticles (NPs) in various matrices. However, ion background concentration is usually much higher in magnitude than NP concentration, causing ionic signals to dominate the mass spectra and mask particle-derived signals, which in turn leads to biased particle size estimation or underestimation of particle number concentrations. Here, we developed a mask-guided reconstruction (MGR) deep learning model for accurate quantification of particle sizes and numbers. This MGR model exhibited remarkable performance on the size and number estimation of NPs compared with traditional noising method, such as wavelet thresholding. Our results indicated that even 1 μg/L free ions would greatly reduce 78.6% of detected NP numbers and lead to the overestimation of the NP size. The MGR model could accurately estimate particle numbers and diameters of Au NPs (>30 nm) in the presence of background ions, which could be further applied to other metal elements without retraining. Furthermore, the MGR model well preserves the contribution of smaller NPs in the biological samples. Overall, this work provides a convenient method to precisely determine NP size and numbers without extra instrument support.