Evaluation of classification methods for identifying multiwalled carbon nanotubes collected on mixed cellulose ester filter media. (9th May 2021)

• Record Type: Journal Article
• Authors: Smith, Devin; Neu‐Baker, Nicole M.; Eastlake, Adrienne C.; Zurbenko, Igor G.; Brenner, Sara A.
• Is Part Of: Journal of microscopy. Volume 283:Part 2(2021)
• Page Start: 102
• Page End: 116
• Publication Date: 2021-05-09
• Subjects: hyperspectral imaging -- microscopy -- occupational exposure assessment -- predictive modelling; Microscopy -- Periodicals; 502.82
• DOI: 10.1111/jmi.13012
• Languages: English
• ISSNs: 0022-2720

Abstract

Abstract: Enhanced darkfield microscopy (EDFM) and hyperspectral imaging (HSI) are being evaluated as a potential rapid screening modality to reduce the time‐to‐knowledge for direct visualisation and analysis of filter media used to sample nanoparticulate from work environments, as compared to the current analytical gold standard of transmission electron microscopy (TEM). Here, we compare accuracy, specificity, and sensitivity of several hyperspectral classification models and data preprocessing techniques to determine how to most effectively identify multiwalled carbon nanotubes (MWCNTs) in hyperspectral images. Several classification schemes were identified that are capable of classifying pixels as MWCNT(+) or MWCNT(–) in hyperspectral images with specificity and sensitivity over 99% on the test dataset. Functional principal component analysis (FPCA) was identified as an appropriate data preprocessing technique, testing optimally when coupled with a quadratic discriminant analysis (QDA) model with forward stepwise variable selection and with a support vector machines (SVM) model. The success of these methods suggests that EDFM‐HSI may be reliably employed to assess filter media exposed to MWCNTs. Future work will evaluate the ability of EDFM‐HSI to quantify MWCNTs collected on filter media using this classification algorithm framework using the best‐performing model identified here – quadratic discriminant analysis with forward stepwise selection on functional principal component data – on an expanded sample set.