Label-free characterisation of sub-wavelength dielectric particles both inside and outside living cells using holographic microscopy and convolutional neural networks

Session

LSA.1 - Label-free life science imaging

Authors

PhD student Erik Olsén (2), PhD, Senior Lecturer Daniel Midtvedt (3), MSc student Benjamin Midtvedt (2, 3), PhD, Post doc Emelie Vilhelmsson Wesén (1), PhD student Fredrik Eklund (2), PhD, Associate Professor Elin Esbjörner (1), PhD, Professor Fredrik Höök (2)

Affiliations

1. Department of Biology and Biological Engineering, Chalmers University of Technology
2. Department of Physics, Chalmers University of Technology
3. Department of Physics, University of Gothenburg

Keywords

convolutional neural networks, dielectric particles, holography, optical properties, particle tracking, subwavelength particles

Abstract text

Characterisation of particles with dimensions smaller than the wavelength of visible light are essential in many fields, ranging from diagnostic and pharmaceutical applications to characterisation of industrial waste products. As both particle size and composition greatly influence particle function, fast and accurate characterisation of these properties has received a growing interest during past decades. However, characterization of dispersed particles in complex local environments is a major challenge as traditional approaches such as nanoparticle tracking analysis (NTA) require particles to be freely diffusing in a medium with known viscosity. This challenge becomes even more prominent for particles internalised by living cells, where the particles motion is restricted and the cellular properties may change over time, making also signal quantification over time difficult. In addition, when investigating properties of living cells, labeling approaches used to visualize subwavelength particles introduce the risk of disturbing the cells’ native state, which in turn limits the applicability of many measurement techniques.

 

To minimize potential sample perturbations, label-free interferometric based approaches such as digital holographic microscopy have therefore gained increased interest due to its ease of use and quantitative information content. By using optical holography, one can for example determine both size and refractive index of freely diffusing dielectric particles down to around 300-400 nm in diameter [1], in additional to quantifying cellular properties as function of cellular state and local environment [2]. However, fast and accurate characterisation of sub-wavelength sized particles using optical holography remains a major challenge, especially in complex sample environments. For example, characterisation based to the Brownian motion require more than 100 independent observations (frames) to generate an accurate size determination on the individual particle level. Furthermore, in the context of particles inside cells, identifying point spread limited particles and quantifying the label-free signal from individual particles require robust and reliable image analysis tools to distinguish the signal of interest from the complex background.

 

To improve the speed and accuracy of particle analysis using optical holography, we use convolutional neural networks for both particle identification and characterisation, where all particle characteristics are extracted directly from the recorded images. By using a weighted average convolutional neural network on a particle mix consisting of 420 nm silica, 300 nm polystyrene and 460 nm polystyrene spheres, all subpopulations could be correctly experimentally identified using only five observations per particle [3]. In addition, both size and refractive index could be accurately determined for all particles with a significantly lower uncertainty compared with holographic NTA using the same number of observations, where this neural network approach require less than an order of magnitude fewer observations to get the same single particle accuracy as ordinary holographic NTA.

 

To demonstrate the potential of this method to characterise subwavelength particles in complex surroundings, fluorescent 450 nm polystyrene particles was internalised in SH-SY5Y cells. When comparing the particles identified by the network with the fluorescent signals from the particles, the network correctly identified almost all particles with only a few percent of false detections as the particles and their optical properties was tracked over time. To compare the characterisation of internalised particles with particles outside the cells, the integrated phase shift, which is proportional to the particle volume times the refractive index difference compared to its local surrounding, was analysed. The determination accuracy of the integrated phase shift was similar for the two cases, taking a minor signal offset corresponding to refractive index of the cytoplasm inside the cells into account. This accuracy opens up the potential to in a label-free manner investigate both nanoparticle uptake across the outer cellular membrane and degradation over time of when internalised into the cell.

 

These results demonstrate further advancement of holographic microscopy aided by implementing neural networks for improved data analysis, which opens up the potential use of the method for investigations of samples which require fast sample characterisation and in complex environments. As optical holography is a label-free imaging technique, minimal sample preparation is needed and by using convolutional neural networks the data is quickly analysed. Further development of this methodological approach may enable optical holography to emerge as an important complementary technique to more wide-spread used fluorescence microscopy approaches, with the potential of providing new insights with respect to both particle and cellular properties and sample dynamics with minimal risk of unwanted sample perturbations.

References

[1] Midtvedt, D. et al., ‘’Size and Refractive Index determination of Sub-Wavelength Particles and Air Bubbles by Holographic Nanoparticle Tracking Analysis’’. Anal. Chem. (2020), 92, 2, 1908-1915

[2] Midtvedt, D. et al., ‘’Label-free spatio-temporal monitoring of cytosolic mass, osmolarity, and volume in living cells’’. Nat Commun (2019), 10, 340.

[3] Midtvedt, B. et al., ‘’ Holographic characterisation of subwavelength particles enhanced by deep learning‘’. In manuscript