Comparison of hyperspectral classification methods for the analysis of cerium oxide nanoparticles in histological and aqueous samples

Hyperspectral imaging (HSI) and classification are established methods that are being applied in new ways to the analysis of nanoscale materials in a variety of matrices. Typically, enhanced darkfield microscopy (EDFM)‐based HSI data (also known as image datacubes) are collected in the wavelength range of 400–1000 nm for each pixel in a datacube. Utilising different spectral library (SL) creation methods, spectra from pixels in the datacube corresponding to known materials can be collected into reference spectral libraries (RSLs), which can be used to classify materials in datacubes of experimental samples using existing classification algorithms. In this study, EDFM‐HSI was used to visualise and analyse industrial cerium oxide (CeO₂; ceria) nanoparticles (NPs) in rat lung tissues and in aqueous suspension. Rats were exposed to ceria NPs via inhalation, mimicking potential real‐world occupational exposures. The lung tissues were histologically prepared: some tissues were stained with hematoxylin and eosin (H&E) and some were left unstained. The goal of this study was to determine how HSI and classification results for ceria NPs were influenced by (1) the use of different RSL creation and classification methods and (2) the application of those methods to samples in different matrices (stained tissue, unstained tissue, or aqueous solution). Three different RSL creation methods – particle filtering (PF), manual selection, and spectral hourglass wizard (SHW) – were utilised to create the RSLs of known materials in unstained and stained tissue, and aqueous suspensions, which were then used to classify the NPs in the different matrices. Two classification algorithms – spectral angle mapper (SAM) and spectral feature fitting (SFF) – were utilised to determine the presence or absence of ceria NPs in each sample. The results from the classification algorithms were compared to determine how each influenced the classification results for samples in different matrices. The results showed that sample matrix and sample preparation significantly influenced the NP classification thresholds in the complex matrices. Moreover, considerable differences were observed in the classification results when utilising each RSL creation and classification method for each type of sample. Results from this study illustrate the importance of appropriately selecting HSI algorithms based on specific material and matrix characteristics in order to obtain optimal classification results. As HSI is increasingly utilised for NP characterisation for clinical, environmental and health and safety applications, this investigation is important for further refining HSI protocols while ensuring appropriate data collection and analysis.