Journal of Quantitative Spectroscopy and Radiative Transfer
Particle sizing by multiangle light-scattering data using the high-order neural networks
Introduction
The problem of retrieval of particle parameters by angular distribution of light scattered by a single particle is actual for different applications, in particular for flow cytometry [1], [2], [3], [4]. Especially, this problem is urgent for the scanning flow cytometry [4]. Nonspherical and spherical particles are the objects of investigation by this method.
There are many approaches to solve this problem. In paper [5], the solution of the inverse problem of scattering for a homogeneous spherical particle is proposed. To determine parameters of a particle, it is necessary to calculate integrals from the measured angular distribution of scattered light multiplied on oscillating functions of the angle in limits from 0° to180°.
The trial-and-error method is widely used [6]. It consists of a multiple solution of a direct problem and selection of particle parameters so that the calculation results as much as possible coincided with the experimentally measured angular distribution of scattered light. The method requires considerable time even for modern computers. To decrease the time of analysis, the empirical techniques for solving an inverse problem of scattering in flow cytometry were offered [4], [7], [8], [9].
Recently, for a solution of the inverse problem of scattering by spherical and nonspherical particles, the neural network method is used [10], [11], [12], [13], [14]. The neural network presents a set of elementary calculators called neurons [15], [16]. Depending on a solved problem, the neurons fulfill different operations. For example, neuron-summator adds the input signals and converts the output signal, according to nonlinear function f(x). The last is a neuron activation function. Usually the S-shaped function of neuron activation is applied. The single-level and multilevel neural networks are used [15], [16]. The arbitrary functions of many variables with a required accuracy can be approximated by the multilevel neural networks of such type [17], [18], [19], [20]. The neural networks, where the neurons fulfill more complicate operations, are applied. For example, in neural networks with the radial basis function (RBF networks), the output signal exponentially decreases with the distance in the Machalanobis metric. Such neural networks are widely used in the pattern-recognition theory. RBF networks are optimal for approximation of functions entered with errors [21].
To solve correctly the problem by neural networks, it is necessary to choose neural network coefficients adequately. This problem is solved on a level of training of the neural network. As a rule, the training of the neural network demands a lot of time. It consists of determining a global extremum of function of many variables. It is a very complicate problem. Due to simplicity and possibility to implement parallel calculations fulfilled by neurons of the network, the trained neural network solves the problem very fastly.
The neural networks with radial basis function are used for identification of particles in flow cytometry [22], [23]. In papers [10], [11], two-level neural networks with radial basis function were proposed to determine parameters of nonabsorbing particles at the range of radius from 0.5 to 1.5 μm and relative refractive index in the range from 1.12 to 1.29. For retrieval of particle parameters, it is necessary to know intensity of scattered light in the range from 0° to 180°.
The neural networks with sigmoidal function of neuron activation were used for retrieval of a mean size of particles and variance of size distribution function [12]. Multilevel neural networks with a symmetric function of neuron activation were applied to determine characteristics of spherical absorbing particles with radius from 0.5 to 5.0 μm by intensity of scattered light in the range from 0° up to 180° [13]. Such neural networks were used for retrieval of volume, refractive index, aspect ratio and orientation of spheroidal particles [14].
The experimental equipment usually allows to measure intensity of scattered light in a limited range of angles. In the scanning flow cytometer, characteristics of scattered light can be measured in the range of angles from 10° to 60°–70° [4], [24].
In the present paper, the problem of retrieval of size and refractive index of spherical particle by intensity of scattered light in the range of angles from 10° to 60° is considered. The ranges of the radius and the refractive index are 0.6–10.6 μm and 1.02–1.38, respectively. In these regions, there are values of parameters of many biological cells. Many of the cells have more complex structure. Their shape cannot be always considered as spherical. Some of them, for example lymphocytes, contain eccentric nuclei. The creation of neural networks for different types of cells has to take into account the specificity of the cell. We consider here only spherical cells in a wide range of optical parameters. This is a step for further investigation of more complex cells.
Section snippets
Forming and training of a neural network
Let us consider a nonabsorbing spherical particle with a relative refractive index n and radius R. The refractive index n varies in the range from 1.02 to 1.38, radius varies in the range from 0.6 to 10.6 μm. The particle is illuminated by a plane wave along z-axis. It is necessary to retrieve R and n by intensity of light , scattered in the directions θj (, Nθ is the amount of angles). Angle θ varies in the range from 10° to 60°. To model scattered light, the Mie theory is
Neural network testing
To test the neural network, we multiply retrieved particle parameters at the presence of the measuring error. A multiplicative error of measurement was modeled in the same manner as described in the previous section. We multiply determined particle parameters on different examples, and by the obtained results, we calculated the root-mean-square deviation of the retrieval error of radius σR and refractive index σn.
Data of Fig. 8 illustrate sensitivity of the single-level and two-level neural
Conclusion
High-order neural networks for retrieval of radius R in the range from 0.6 to 10.6 microns and refractive index n in the range from 1.02 to 1.38 of spherical nonabsorbing particles by the relative intensity of scattered light in the range of angles from 10° to 60° are developed.
The possibilities of retrieval of particle parameters by multilevel neural networks are analyzed at changes in the amount of neurons and other structural parameters of the neural network.
Retrieval errors of R and n are
Acknowledgement
This research is partially sponsored by NATO's Scientific Affairs Division in the framework of the Science for Peace Programme.
References (29)
- et al.
Multilayer feedforward networks are universal approximators
Neural Networks
(1989) - et al.
Sizing of soft spheroidal particles by multiangle scattered light intensity data: application of neural network
JQSRT
(2004) - et al.
Sizing of spheroidal and cylindrical particles in a binary mixture by measurement scattered light intensity data: application of neural network
JQSRT
(2005) Practical flow cytometry
(1995)Flow cytometry: a practical approach
(2000)- et al.
Biophysical and biomedical applications of nonspherical scattering
Scanning flow cytometry for individual particle analysis
Rev Sci Instrum
(2000)- et al.
Inverse Mie problem
J Opt Soc Am A
(2000) - et al.
Flow-cytometric light scattering measurement of red blood cell volume and hemoglobin concentration
Appl Opt
(1985) - et al.
Calibration—free method to determine the size and hemoglobin concentration of individual red blood cells from light scattering
Appl Opt
(2000)
Single-particle sizing from light scattering by spectral decomposition
Appl Opt
Size and refractive index of individual marine particulates: a flow cytometric approach
Appl Opt
Application of neural networks to the inverse light scattering problem for spheres
Appl Opt
On solving the inverse scattering problem with RBF neural networks: noise-free case
Neural Comput Appl
Cited by (10)
GPU-accelerated repulsive particle swarm optimization for parameter retrieval of single sphere from angularly distributed scattered light
2020, OptikCitation Excerpt :For example, radial basis function neural networks were employed to inversely model the scattering by a sphere in Ref. [12]. Berdnik V. et al. used high-order neural networks to perform particle sizing with multi-angle light-scattering data [13]. Characterization of hexagonal prisms based on 2D scattering patterns was performed by machine learning [14].
WEBSCAT: A web application for the analysis of electromagnetic scattering from small particles
2014, Journal of Quantitative Spectroscopy and Radiative TransferCitation Excerpt :This angular variation of the scattered signal also depends on the physical (size, shape, porosity etc.) as well as optical (refractive index) properties of the scatterer and therefore carries its signature. Inversely it is possible to utilize these signatures or scattering informations to extract the characteristics of the scatterer [18–21]. Thus study of the angular scattering dependency of small particulate matter not only helps for better understanding of radiation transfer through a medium containing the scatterer but is a unique tool to classify or even identify the scattering particle [22].
A statistical complement to deterministic algorithms for the retrieval of aerosol optical thickness from radiance data
2006, Engineering Applications of Artificial IntelligenceNeural networks for particle parameter retrieval by multi-angle light scattering
2015, Light Scattering Reviews 10: Light Scattering and Radiative TransferLight scattering and morphology of the lymphocyte as applied to flow cytometry for distinguishing healthy and infected individuals
2010, Journal of Biomedical OpticsSizing homogeneous spherical particles from intensity-only angular scatter
2010, Journal of the Optical Society of America A: Optics and Image Science, and Vision