rus
Issue #5/2021
O.V.Sinitsyna, M.M.Vorob’ev, I.V.Yaminskiy
Automated search for nanoparticles in the probe microscope images using a neural network
Automated search for nanoparticles in the probe microscope images using a neural network
DOI: 10.22184/1993-8578.2021.14.5.276.280
There are a number of important advantages that can be obtained by using automated search for objects detected by the probe microscopy, such as a high rate of data processing, minimization of experimentalist’s influence on the measurement process and a possibility to enlarge the data volume to be processed. It has been shown in this work that for the search for nanoparticles, which dimensions are comparable to a noise level, the more accurate result is achieved by a neural network algorithm based on protein nanoparticles data processing which was obtained using the atomic force microscopy while the larger nanoparticles are more precisely distinguishable by the threshold algorithm.
There are a number of important advantages that can be obtained by using automated search for objects detected by the probe microscopy, such as a high rate of data processing, minimization of experimentalist’s influence on the measurement process and a possibility to enlarge the data volume to be processed. It has been shown in this work that for the search for nanoparticles, which dimensions are comparable to a noise level, the more accurate result is achieved by a neural network algorithm based on protein nanoparticles data processing which was obtained using the atomic force microscopy while the larger nanoparticles are more precisely distinguishable by the threshold algorithm.
Теги: atomic-force microscopy nanoparticles neural network scanning probe microscopy атомно-силовая микроскопия наночастицы нейронная сеть сканирующая зондовая микроскопия
Automated search for nanoparticles in the probe microscope images using a neural network
INTRODUCTION
One of the main methods to study the nanooblects is a scanning probe microscopy (SPM). The basic advantages of SPM are: a high spatial resolution, the ability to construct a three-dimensional surface map, and receipt of information about the local physical and chemical properties of the surface and fixed objects on it [1]. Implementation of a concept that uses artificial intelligence makes it possible to get more information received by SPM method, and reduce the experimentalist’s influence on the data processing which can lead to significant distortion of the results even when making routine measurements [2].
Currently the artificial intelligence algorithms are quite rarely used in SPM [3, 4]. One of the main difficulties is the need to collect massive experimental data to test the algorithms. Artificial intelligence algorithms can be useful in SPM to search for the objects in the images and acquisition of their quantitative information of the microstructure. Simple threshold searching algorithms in the case of small objects usually produce insufficient results.
RESEARCH METHODS
In this work, we applied the data processing algorithm using a neural network to search for the protein nanoparticles in the images obtained by the atomic force microscopy (AFM), and analyze their size distribution. Use was made of the FemtoScan Online software [5] for AFM data processing and isolation of the objects applying the threshold algorithm and Gwyddion [6] to search for the objects with the aid of a neuron network.
An AFM image of the protein nanoparticles applied onto the mica surface from the water solution and dried in air [7] is shown in Fig.1. The image was obtained with the aid of a FemtoScan microscope (Advanced Technologies Center) in a semi-contact scanning mode. The particle height is comparable to a noise level, which is 0.3 nm, as it was seen on the indicated surface profile. We suppose that in this case the noise level is dependent on the low-molecular components deposited onto mica from the solution.
The results of a comparison of manual approach and automated methods of protein nanoparticles sizes detection are presented in Fig.2. Analyzed are the maximum height of the particles and disk radius of the area that is equivalent to the area occupied by one particle. The neural network, realized in Gwyddion software, was used to eliminate the background between particles, thereafter the threshold algorithm was applied. In the manual approach, contours of each particle were described by an ellipse.
If a noise level is comparable with the size of the particles, application of the threshold method leads to significant distortion of the particle size distribution in the small heights and radii area because of wrong detection of the surface fields, which do not contain nanoparticles, or due to incorrect isolation of the contours of the objects. Use of the neural network to eliminate the background can significantly reduce the number of "false" particles and construct contours more precisely. However, in the big heights area, where the particle sizes are much greater than the noise level, the threshold method yields more accurate results. It can be due to the fact that just one-third number of the particles has a height of more than 1 nm, and distribution of sizes for them is extremely inhomogeneous, therefore, application of the neural network is not reasonable because of the number of big particles is not enough for its training. It is necessary to point out a downgrading of the radius of the disk having the equivalent area in case when the automated methods for isolation of the nanoparticle contours are used (see Fig.2b).
CONCLUSIONS
The probe microscopy methods ensure a high spatial resolution but in many cases only a relatively small surface area is observed, and a large data array is not possible due to the wear of the probe or time limitations.
The results of the work show that in such a common practical situation, when it is required to upgrade quality of the automatic search for nanoparticles, it is necessary to use a combination of a threshold algorithm for large particles and a neural network for those particles which dimensions are comparable to noise level.
ACKNOWLEDGEMENTS
The study was completed with the financial support of the RFBR (grant No. 21-58-10005) and Ministry of Science and Higher Education of the Russian Federation. ■
Declaration of Competing Interest. The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
INTRODUCTION
One of the main methods to study the nanooblects is a scanning probe microscopy (SPM). The basic advantages of SPM are: a high spatial resolution, the ability to construct a three-dimensional surface map, and receipt of information about the local physical and chemical properties of the surface and fixed objects on it [1]. Implementation of a concept that uses artificial intelligence makes it possible to get more information received by SPM method, and reduce the experimentalist’s influence on the data processing which can lead to significant distortion of the results even when making routine measurements [2].
Currently the artificial intelligence algorithms are quite rarely used in SPM [3, 4]. One of the main difficulties is the need to collect massive experimental data to test the algorithms. Artificial intelligence algorithms can be useful in SPM to search for the objects in the images and acquisition of their quantitative information of the microstructure. Simple threshold searching algorithms in the case of small objects usually produce insufficient results.
RESEARCH METHODS
In this work, we applied the data processing algorithm using a neural network to search for the protein nanoparticles in the images obtained by the atomic force microscopy (AFM), and analyze their size distribution. Use was made of the FemtoScan Online software [5] for AFM data processing and isolation of the objects applying the threshold algorithm and Gwyddion [6] to search for the objects with the aid of a neuron network.
An AFM image of the protein nanoparticles applied onto the mica surface from the water solution and dried in air [7] is shown in Fig.1. The image was obtained with the aid of a FemtoScan microscope (Advanced Technologies Center) in a semi-contact scanning mode. The particle height is comparable to a noise level, which is 0.3 nm, as it was seen on the indicated surface profile. We suppose that in this case the noise level is dependent on the low-molecular components deposited onto mica from the solution.
The results of a comparison of manual approach and automated methods of protein nanoparticles sizes detection are presented in Fig.2. Analyzed are the maximum height of the particles and disk radius of the area that is equivalent to the area occupied by one particle. The neural network, realized in Gwyddion software, was used to eliminate the background between particles, thereafter the threshold algorithm was applied. In the manual approach, contours of each particle were described by an ellipse.
If a noise level is comparable with the size of the particles, application of the threshold method leads to significant distortion of the particle size distribution in the small heights and radii area because of wrong detection of the surface fields, which do not contain nanoparticles, or due to incorrect isolation of the contours of the objects. Use of the neural network to eliminate the background can significantly reduce the number of "false" particles and construct contours more precisely. However, in the big heights area, where the particle sizes are much greater than the noise level, the threshold method yields more accurate results. It can be due to the fact that just one-third number of the particles has a height of more than 1 nm, and distribution of sizes for them is extremely inhomogeneous, therefore, application of the neural network is not reasonable because of the number of big particles is not enough for its training. It is necessary to point out a downgrading of the radius of the disk having the equivalent area in case when the automated methods for isolation of the nanoparticle contours are used (see Fig.2b).
CONCLUSIONS
The probe microscopy methods ensure a high spatial resolution but in many cases only a relatively small surface area is observed, and a large data array is not possible due to the wear of the probe or time limitations.
The results of the work show that in such a common practical situation, when it is required to upgrade quality of the automatic search for nanoparticles, it is necessary to use a combination of a threshold algorithm for large particles and a neural network for those particles which dimensions are comparable to noise level.
ACKNOWLEDGEMENTS
The study was completed with the financial support of the RFBR (grant No. 21-58-10005) and Ministry of Science and Higher Education of the Russian Federation. ■
Declaration of Competing Interest. The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Readers feedback
