Open Access
Add to BibSonomy
Abstract
Exosomes are extracellular vesicles that serve as promising intrinsic nanoscale biomarkers for disease diagnosis and treatment. Nanoparticle analysis technology is widely used in the field of exosome study. However, the common particle analysis methods are usually complex, subjective, and not robust. Here, we develop a three-dimensional (3D) deep regression-based light scattering imaging system for nanoscale particle analysis. Our system solves the problem of object focusing in common methods and acquires light scattering images of label-free nanoparticles as small as 41 nm in diameter. We develop a new method for nanoparticle sizing with 3D deep regression, where the 3D time series Brownian motion data of single nanoparticles are input as a whole, and sizes are output automatically for both entangled and untangled nanoparticles. Exosomes from the normal and cancer liver cell lineage cells are observed and automatically differentiated by our system. The 3D deep regression-based light scattering imaging system is expected to be widely used in the field of nanoparticle analysis and nanomedicine.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Exosomes are a type of extracellular vesicles with relatively small size, usually between 40-150 nm [1–3]. The role of exosomes in the development of diseases has been studied extensively [4,5]. Nanoscale exosomes can easily penetrate the interstitium of organs and tumors [6], with a great potential for drug delivery. The small exosomes can also pass through impermeable biological barriers such as the blood-brain barrier [7]. Compared with cell-level objects such as circulating tumor cells, exosomes are distributed in almost all body fluids, which may be widely used in liquid biopsy [8]. The bioactive substances contained in exosomes are expected to realize skin anti-aging, thus exosomes can be applied to the field of medical cosmetology [9]. These demonstrate the wide applications of exosomes in many fields, however the nanoscale size leads to great challenges for accurately quantifying and localizing single exosomes.
Current methods for exosome detection mainly include transmission electron microscopy (TEM), atomic force microscopy (AFM), nano-flow cytometry, and nanoparticle tracking analysis (NTA) technique. Transmission electron microscopy and AFM are the detection technologies that can reflect the physical information of exosomes [10,11], but have the disadvantages such as complex operation and high cost. Nano-flow cytometry and imaging flow cytometry are the recently emerged methods for exosome detection [12,13]. Both of them can detect single exosome with high throughput, but they usually require fluorescence labeling for effective measurements. Nanoparticle tracking analysis is the most commonly used exosome analysis technology [14,15], which is relatively simple for implementations in optical or chemical laboratories.
For nanoparticle analysis in NTA, the Brownian motion trajectories of the exosomes are collected, then the particle sizes are calculated by trajectory analysis. Recently, Moore et al. have developed a method called multispectral nanoparticle tracking analysis, which utilizes three lasers of different wavelengths to simultaneously excite nanoparticles of polydisperse samples [16]. Dai et al. have combined Raman spectroscopy detection with NTA, which can simultaneously detect the Raman signal and Brownian motion of exosomes, and an additional exosome image screening process has been included to remove unqualified particle images [17]. Light scattering method obtains the optical information of illuminated objects due to the dipole radiation [18], which is well suited for label-free nanoparticle and cellular analysis. In 2007, Su et al. initiated the 2D light scattering technology for human cell analysis, where the 2D light scattering is correlated with the nanoscale mitochondria [19]. Recently, the 2D light scattering technology has been applied for label-free differentiation of many types of cells by several groups [20–23]. It is interesting to explore the 2D light scattering technology for the analysis of label-free nanoscale exosomes instead of single cells.
In NTA, the randomness of Brownian motion and the uncertainty of data analysis may cause certain obstacles to more accurate analysis of exosomes. Due to the factors such as illumination, background noise, and the complex point spread function in data acquisition, particle localization becomes difficult and subjective. In recent years, new algorithms such as radial basis function and deep learning methods have been adopted to improve the accuracy of particle localization [24–26]. Recently, Wang et al. have designed the deepEVanalyzer for the detection of exosomes, which combines light-sheet illumination and deep learning to improve the analysis results [27]. Above studies demonstrate that researchers have adopted deep learning or other machine leaning techniques for nanoparticle identification or localization, but the size calculation based on nanoparticle Brownian motion by using intelligent algorithms has not been well studied. Unlike networks for classification tasks [28,29], deep regression networks have advantages when making predictions for continuous data with high data analysis capabilities, such as DeepTrack [26] and vgg16–regression [30] for 2D data input. Moreover, the 3D regression network has obvious advantages in data processing [31], which is worthy to be explored for the analysis of the nanoparticle 3D time series data. Secondly, the parameters used in the mean square displacement fitting are controversial while using different fitting parameters for the diffusion coefficient calculation [32–34]. Thirdly, since the Brownian motion of nanoparticles is random, the number of tracking frames is also an important parameter affecting the analysis accuracy. Conventionally, large number of tracking frames is suggested for nanoparticle size analysis, while small frame number can cause inaccurate size calculations [35–37]. However, nanoparticles with smaller sizes have faster movements, making it impossible to track them for a long time. Therefore, it is of great significance to develop new methods for nanoparticle size analysis to solve the above problems.
In this paper, we develop a 3D deep regression-based light scattering imaging system to enable automated particle size analysis of nanoparticles and exosomes. The light-sheet illumination technique is exploited to acquire light scattering images from nanoparticles. A 3D deep regression network is deveped to intelligently perform single particle size analysis, which can directly obtain the final nanoparticle sizes from the 3D time series image data that are input as a whole. This simplifies the data process procedures for nanoparticle analysis, and avoids the subjective factors in commonly used analysis methods. Results of simulated nanoparticle data, standard nanoparticles, and exosomes of normal and cancerous liver cell lineage cells are obtained and analyzed. Our 3D deep regression-based light scattering imaging system has the advantages of 3D data automatic analysis as a whole and the label-free and high-resolution light scattering imaging, as compared with the common nanoparticle or exosome analysis modalities.
2. Materials and methods
2.1 3D deep regression-based light scattering imaging system
As shown in Fig. 1, our 3D deep regression network-based light scattering imaging system mainly includes light scattering imaging part and 3D deep regression analysis part. The light scattering imaging part adopts light sheet illumination as previously reported [27]. Here a 532 nm laser (Frankfurt, Germany) is used as the excitation light source, and about 14 µm (in thickness) light sheet is formed by a cylindrical lens (Thorlabs, USA), which excites the nanoparticles via an illuminating objective (Olympus, Japan). The sample is injected into a specially-made glass chip (inset in Fig. 1), with a 25 mm x 2 mm x 1 mm size sample room. The scattered light of the label-free nanoparticles is imaged via a 20X objective (NA = 0.4, Olympus, Japan) on a CMOS detector (Canon, Japan). Brownian motion videos of each single nanoparticles are recorded and saved as 3D time series image data. The 3D deep regression analysis part includes a trained deep regression network model, which can automatically obtain the nanoparticle sizes from the 3D time series image data of single nanoparticles that are input as a whole.
Fig. 1. 3D deep regression-based light scattering imaging system. The light scattering imaging part is used to track nanoparticles such as exosomes; the 3D deep regression analysis part is used to perform intelligent single nanoparticle size analysis, where the single-particle 3D time series data are input to the deep learning model that directly outputs the nanoparticle sizes.
2.2 Generation of simulated light scattering images of nanoparticles
In order to train the neural network, we built a training data set using simulated data. Compared with experimental data, simulated data has exact nanoparticle size values and can be acquired with self-defined sizes. The establishment of simulation data mainly includes two parts: one is the Brownian motion simulation of nanoparticles, and the other is the particle image simulation based on point spread function. The Brownian motion simulation assumes that the distance moved within the time interval $\textrm{dT}$. is uncorrelated over time, which follows a normal distribution with a mean distance of zero. In this paper we consider nanoparticles moving in two dimensions, and the motion position of each dimension can be carried out separately and independently. Equations (1) and (2) are used to obtain the position coordinates of nanoparticles in each frame:
Among them, $\textrm{i} = 1,2,3 \ldots \textrm{N} - 1$, represents the number of frames of particle motion, and N is 10 or 20 or 30; ${k_B}$ is the Boltzmann constant; ${k_T}$ represents the absolute temperature; $\textrm{dT}$ is the time interval of each frame (0.02s in this study); η is the coefficient of viscosity; D is the diameter of the particle; $\textrm{GX}$ and GY are two random sequences of length N produced by a normal distribution. Ps represents the side length of each square pixel in the simulated image. x and y represent coordinates in pixels in the image. The length and width of the training images in this paper are both 128 pixels. Here we set ${x_1}$ and ${y_1}$ to random numbers between 51 - 71, $G{X_1}$ and $G{Y_1}$ to 0, so the particles start moving at a random point near the image center. We generate a point spread function based on Fraunhofer diffraction, as shown in Fig. 2(a). Figure 2(b) is a representative image of 50 nm particle obtained by the point spread function. Combined with the simulated Brownian motion trajectory coordinates (Fig. 2(c)), we obtain the 3D data for the tracking of the 50 nm simulated particles in time series (Fig. 2(d)). The simulated data of nanoparticles with different sizes are used to construct a training dataset.
Fig. 2. Simulated data generation of light scattering images of nanoparticles in time series. (a) is a 3D diagram of the point spread function used to generate the nanoparticle image. (b) is a representative 50 nm nanoparticle light scattering image from the point spread function. (c) is a randomly generated 2D Brownian motion trajectory (30 frames). (d) is the illustration of final 3D time series nanoparticle data generated using Brownian motion trajectories and nanoparticle light scattering images.
2.3 3D deep regression model design
Our 3D deep regression network design is shown in Fig. 3. It consists of five 3D convolutional layers, five Batch normalization layers, five rectified linear unit layers, five 3D maximum pooling layers, one dropout layer, a fully connected layer, and a regression output layer. The input is 128 × 128 x N 3D data, where N is the number of frames corresponding to the data of light scattering images. Finally, a value is output from the regression layer, which represents the calculated nanoparticle size. This model structure combines 3D convolution layer and regression task, which has advantages in processing 3D time series data of nanoparticles and in obtaining their continuous size distribution. All the algorithms are carried out on a computer (AMD Ryzen 9 3900X, NVIDIA 2080Ti, 64 GB RAM) with MATLAB 2021a software (Mathworks, USA).
Fig. 3. Schematic diagram of 3D deep regression network structure. Conv-3d: 3D convolutional layer; Batchnorm: Batch normalization layer; Relu: Rectified linear unit layer; Maxpool-3d: 3D maximum pooling layer; Dropout: dropout layer; FC: Fully connected layer.
2.4 Sample preparation
Standard polystyrene beads (Invitrogen, USA) with diameters of 41 nm (SD = 7 nm) and 120 nm (SD = 7 nm) were diluted with ultrapure water and sonicated for 5 min before use. Exosomes were extracted from the cell supernatant of CCC- HEL-1 cells (Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences, Beijing, China) and Hep G2 cells (Procell Life Sciences & Technology Co., Ltd., China). Briefly, the cell supernatant treated with exosome-free medium was collected and concentrated by centrifugation. An exosome separation column (Izon, New Zealand) was used then to collect exosomes.
3. Results and discussion
3.1 Light scattering imaging of 41 nm nanoparticles
The light scattering imaging part is the basis for deep regression analysis. In our system, the light sheet technology can avoid adding too much background noise and reduce the influence of defocusing on the nanoparticles. Standard polystyrene beads with a diameter of 41 nm were used to test the ability of the imaging part, as this size is the lower limit of the exosome size distribution. The experimental results are shown in Fig. 4. Each image collected contains multiple particles. As can be seen from the full image and zoom-in images of three typical particles, our light scattering imaging system can clearly detect nanoparticle with diameters as low as 41 nm.
Fig. 4. Representative 41 nm nanoparticle light scattering images obtained by our system. The top row is the full-size images from a 3D time series data that we obtained from the light scattering imaging part, image size: 640 × 240 pixels. The bottom row is the zoom-in view of three single nanoparticles.
3.2 Validation of the 3D deep regression-based light scattering imaging system with simulated nanoparticles
We construct a deep learning training data set using simulated data of light scattering images as described in the section of Materials and Methods. In order to make our network focus on exosome size, the diameter of each training particle is randomly generated in the range of 30 - 200 nm with a step size of 10 nm when constructing the training dataset. In this paper, 10, 20, and 30 frame lengths deep regression models are trained and the corresponding training dataset sizes are 20000, 10000 and 7000 particles, respectively. In order to further eliminate the influence caused by random training, we trained 5 replicate models with the same conditions and training data set, and took the average of the calculated results as the final particle size. To verify the capability of our 3D deep-regression model, we simulated the Brownian motion data of nanoparticle light scattering images with a fixed size and a fixed number of frames. The simulation data of 50 nm, 80 nm, 120 nm, and 150 nm nanoparticles are used for testing. Three sets of test data are constructed for each size, and each set of test data contains 1000 particles. It should be noted that the simulated data used for testing does not overlap with the training data.
The mean value, modal value, and coefficient of variation (CV) of each group of nanoparticles were calculated by our deep regression method (DRM), and compared with the common method (CM) and the maximum likelihood estimation correction method (MLECM), as shown in Table 1. The mean value represents the arithmetic mean of all the particle size results. The modal values represent the values that appear most frequently in all the particle size results. The coefficient of variation is defined as the ratio of standard deviation to the mean value, which represents the dispersion degree of the particle size distribution. The CM is obtained by the ImageJ plug-in “Nanotrackj” [34]. Here the maximum value method is used for particle positioning, the mean square displacement regression method is used for diffusion coefficient calculation, and the particle size is calculated by the Stokes Einstein equations. The MLECM uses the maximum likelihood estimation to correct the particle size distribution, and obtains a histogram of particle size distribution closer to the true distribution [34,38]. It can be seen from Table 1 that both methods (CM and MLECM) can obtain reliable results in the case of 30 frame steps, while our DRM results are much closer to the real values. We also notice that as the number of nanoparticle frames decreases, the obtained particle size analysis results are more diverged. However, our DRM is more robust than the other two methods, when the number of frames varies. It can be seen that both the CM and the MLECM can obtain acceptable results at a length of 30 frames, but the results of 10 frames are unreliable. Here the CV of the DRM is less than that of the CM and the MLECM for every test set, which indicates that our method offers more concentrated and narrower results for particle size distribution. In addition, as the particle size increases, the results from the CM and the MLECM become worse as compared with DRM. This may be due to that the slower Brownian motion of larger nanoparticles leads to bigger deviation in displacement calculation, which affects the results of CM and MLECM, but our method is barely affected by this effect.
Table 1. Comparison of nanoparticle size analysis results of our deep regression method (DRM), common method (CM) and maximum likelihood estimation correction method (MLECM) for the simulated data set.a
3.3 3D deep-regression light scattering imaging system for the analysis of 41 and 120 nm particles
So far, we have demonstrated the capability of our deep regression model for simulated data, as shown in Fig. 5 (a). In experimental light scattering imaging data, due to the different sample concentrations, there may be more than one particle in the region of interest. We call them untangled particle (Fig. 5 (b)) or entangled particles (Fig. 5 (c)), which represents the cases of one single particle or multiple particles (3, 5, 11 particles here as representative results), respectively. For untangled particle, only simple filtering and thresholding are needed to deal with the background noise, which can be directly input into the deep regression network. For the entangled particles, simple preprocessing to split them into single particle motion data is needed to meet the conditions of the deep regression model. First, we use the maximum method to locate each point in the image and then use the plug-in “TrackMate” of ImageJ to track each particle so as to obtain the motion trajectory coordinates [39], as shown in Fig. 5 (d). After obtaining the trajectory of each particle, the image can be cropped to 128 × 128 x N 3D data according to the first position of each trajectory. It should be noted that this motion trajectory is not used to calculate the particle size, but is only used to split the mixed particles. After splitting, Gaussian filtering and threshold segmentation are used to remove the background noise and the extra particles to make it more similar to the training data. Finally, the single particle 3D time series data is obtained as shown in Fig. 5 (e), which was input into the trained neural network model to obtain the nanoparticle size.
Fig. 5. Display of 3D time series data and preprocessing of the entangled particles. (a) is the simulated data used for training. (b) is the experimental untangled particle, and there is only one particle in a 3D time series data. (c) shows the experimental entangled particles data before preprocessing. (d) shows the trajectories of the five particles in the middle image of (c). (e) shows the single particle time series data that has been decomposed after preprocessing.
To test the effectiveness of our DRM, 41 nm and 120 nm polystyrene nanoparticles were analyzed. Due to the faster Brownian motion of small particles, we found that most of the trajectory lengths are concentrated below 40 frames. After removing the trajectories with less than 10 frames [40,41], it is found that more than 75% of the trajectories of the 41 nm beads are less than 40 frames. Even for the data of 120 nm particles whose Brownian motion is slower than 41 nm particles, the percentage is still more than 55%. This means that the accuracy of short track analysis is crucial. Therefore, we use the trained three different deep regression models with different frames (10 frames, 20 frames, 30 frames) to examine the particle size calculation effect of experimental samples under 40 frames. That is, using the 10 frames model, the 20 frames model, and the 30 frames model to calculate the particle trajectories between 10 - 20 frames, 20 - 30 frames, and 30 - 40 frames, respectively. The final size cumulative distribution results are shown in Fig. 6. It can be seen that our method (blue dashed line) is closer to the gold standard (black solid line) than the common method (orange dash-dotted line) and the maximum likelihood estimation correction method (green dotted line). A total of 258 particles (41 nm sample) and 176 particles (120 nm sample) were analyzed, and the average diameter of the two types of nanoparticles calculated by our method is 46.5 nm (SD = 17.4 nm) and 119 nm (SD = 17.4 nm), which is close to the true value. However, the values obtained with CM are 67 nm (SD = 56.9 nm) and 166.5 nm (SD = 107.9 nm), respectively. The results obtained with MLECM are 59.3 nm (SD = 45 nm) and 152 nm (SD = 84.2 nm), respectively, which are better than the common method, but are inferior to our 3D deep regression method.
Fig. 6. Comparison of cumulative distribution results for experimental particle data. (a) is the analysis result of 41 nm polystyrene nanoparticles. (b) is the analysis result of 120 nm polystyrene nanoparticles. DRM: our deep regression method; CM: common method; MLECM: maximum likelihood estimation correction method; Gs: the gold standard of the polystyrene nanoparticles.
3.4 Analysis of exosomes by 3D deep regression-based light scattering imaging system
Our 3D deep regression-based light scattering imaging system is applied for the analysis of exosomes from human normal liver cell lineage cells (CCC-HEL-1) and cancerous liver cell lineage cells (Hep G2). Figures 7 (a) and 7 (c) show the TEM photographs of liver cancer cell exosomes and normal liver cell exosomes, respectively. The yellow arrows indicate the typical exosomes. Their light scattering images are shown in Figs. 7 (b) and 7 (d), respectively. It can be seen that our light scattering imaging system can collect a large number of exosomes, and can provide the 3D data of single exosome for nanoparticle size analysis. It is however challenging to obtain many images of single exosome from TEM for statistical analysis, and it is also not practical to obtain the 3D time serious data by TEM.
Fig. 7. Measurements of exosomes from cancerous and normal liver cell lineage cells by our 3D deep regression-based light scattering system. (a) is the TEM photograph of liver cancer cell exosomes. (b) shows the light scattering images of liver cancer cell exosomes. (c) is the TEM photograph of normal liver cell exosomes. (d) shows the light scattering images of normal liver cell exosomes obtained by our system.
After the light scattering imaging measurements and data processing of exosomes, our deep regression model is applied to calculate the exosome size distribution. We analyzed 12 videos of exosomes, 6 from the cancerous liver cells and 6 from the normal ones. A total of 2612 normal liver cell exosomes and 2912 liver cancer cell exosomes were collected and analyzed, the average size of the normal liver cell exosomes and the liver cancer cell exosomes is 98 nm (SD = 30 nm) and 89 nm (SD = 31 nm), with an averaged size difference of 9 nm. Our results show that the averaged size of normal liver cell exosomes is generally larger than that of liver cancer cells, which demonstrates the potential of exosomes as intrinsic markers for disease diagnosis such as cancer detection. The analysis capability of exosomes by our system is also compared with the common method, and the representative analysis results of one of the exosome videos are shown in Fig. 8. As can be seen from Fig. 8 (a), our analysis results are basically in the recognized exosome distribution range of 40-150 nm [1–3], and only about 3.4% of particles exceed 150 nm. For the size of exosomes obtained by common method, at least 29.8% of the particles exceed 150 nm. In addition, it is found that 5.2% of the particles exceed 300 nm in size by the common method. This may be due to the limited performance of the common method for short trajectory data of exosomes. Please note that the last column is for particles with sizes distributed from 300 nm to infinity, while the other columns shown in Fig. 8(b) are for particles with sizes in a 4 nm range of width. Furthermore, the Gaussian fits illustrate that the peaks of the particle size distribution are around 100 nm.
Fig. 8. Exosome size analysis comparison of our 3D deep regression method and the common method. The exosomes are from human cancerous liver cell lineage cells (Hep G2) and normal cells (CCC-HEL-1). (a) is the exosome size distribution obtained by our 3D deep regression network. (b) is the exosome size distribution result obtained by the common method. The solid orange line is for size = 150 nm.
4. Conclusions
In summary, we developed the first generation of an integrated 3D deep regression network based light scattering imaging system for the tracking and analysis the size of exosomes and nanoparticles. Light sheet illumination and light scattering imaging techniques are adopted to obtain the tracking images of nanoparticles without the necessity of tight optical focusing. Different from the common Brownian motion method for nanoparticle sizing, the 3D deep regression network enables direct calculation of the nanoparticle size while using the single particle 3D time series data as a whole. Experimental results show that our system can provide attractive visualization and accurate sizing for nanoparticles with diameters in between 40 and 150 nm, which outperforms the common method and the maximum likelihood estimation correction method. It is also demonstrated that exosomes of human normal and cancer liver cell lineage cells can be differentiated by their size distribution as determined by our 3D deep regression-based light scattering imaging system. To the best of our knowledge, this is probably the first combination of 3D deep regression for label-free nanoparticle size analysis with light scattering imaging, especially for the exosomes.
Funding
Shandong Provincial Key Research and Development Program (Major Scientific and Technological Innovation Project) (2019JZZY011016); Training Program of the Major Research Plan of the National Natural Science Foundation of China (91859114); Fundamental Research Funds for the Central Universities (2022JC025).
Disclosures
The authors declare no conflicts of interest.
Data Availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. Y. Yan, G. Fu, and L. Ming, “Role of exosomes in pancreatic cancer,” Oncol. Lett. 21(4), 7479–7488 (2018). [CrossRef]
2. A. Makler and W. Asghar, “Exosomal biomarkers for cancer diagnosis and patient monitoring,” Expert Rev. Mol. Diagn. 20(4), 387–400 (2020). [CrossRef]
3. M. P. Bebelman, P. Bun, S. Huveneers, G. van Niel, D. M. Pegtel, and F. J. Verweij, “Real-time imaging of multivesicular body–plasma membrane fusion to quantify exosome release from single cells,” Nat. Protoc. 15(1), 102–121 (2020). [CrossRef]
4. Y. L. Tai, K. C. Chen, J. T. Hsieh, and T. L. Shen, “Exosomes in cancer development and clinical applications,” Cancer Sci. 109(8), 2364–2374 (2018). [CrossRef]
5. K. Stefanius, K. Servage, and K. Orth, “Exosomes in cancer development,” Curr. Opin. Genet. Dev. 66, 83–92 (2021). [CrossRef]
6. B. Corradetti, D. Gonzalez, I. Mendes Pinto, and R. S. Conlan, “Exosomes as therapeutic systems,” Front. Cell. Dev. Biol. 9, 714743 (2021). [CrossRef]
7. M. Zhang, X. Zang, M. Wang, Z. Li, M. Qiao, H. Hu, and D. Chen, “Exosome-based nanocarriers as bio-inspired and versatile vehicles for drug delivery: recent advances and challenges,” J. Mat. Chem. B 7(15), 2421–2433 (2019). [CrossRef]
8. B. Zhou, K. Xu, X. Zheng, T. Chen, J. Wang, Y. Song, Y. Shao, and S. Zheng, “Application of exosomes as liquid biopsy in clinical diagnosis,” Signal Transduction Targeted Ther. 5(1), 1–14 (2020). [CrossRef]
9. J. Wu, S. Wu, L. Zhang, X. Zhao, Y. Li, Q. Yang, R. Yuan, J. Liu, H. Mao, and N. Zhu, “Stem cell-derived exosomes: a new method for reversing skin aging,” Tissue Eng. Regen. Med. 19(5), 961–968 (2022). [CrossRef]
10. S. Sharma, M. LeClaire, and J. Gimzewski, “Ascent of atomic force microscopy as a nanoanalytical tool for exosomes and other extracellular vesicles,” Nanotechnology 29(13), 132001 (2018). [CrossRef]
11. A. Kotrbová, K. Štěpka, M. Maška, J. J. Pálenik, L. Ilkovics, D. Klemová, M. Kravec, F. Hubatka, Z. Dave, and A. Hampl, “TEM ExosomeAnalyzer: a computer-assisted software tool for quantitative evaluation of extracellular vesicles in transmission electron microscopy images,” J. Extracell. Vesicles 8(1), 1560808 (2019). [CrossRef]
12. F. L. Ricklefs, C. L. Maire, R. Reimer, L. Dührsen, K. Kolbe, M. Holz, E. Schneider, A. Rissiek, A. Babayan, and C. Hille, “Imaging flow cytometry facilitates multiparametric characterization of extracellular vesicles in malignant brain tumours,” J. Extracell. Vesicles 8(1), 1588555 (2019). [CrossRef]
13. Y. Tian, M. Gong, Y. Hu, H. Liu, W. Zhang, M. Zhang, X. Hu, D. Aubert, S. Zhu, and L. Wu, “Quality and efficiency assessment of six extracellular vesicle isolation methods by nano-flow cytometry,” J. Extracell. Vesicles 9(1), 1697028 (2020). [CrossRef]
14. R. A. Dragovic, C. Gardiner, A. S. Brooks, D. S. Tannetta, D. J. Ferguson, P. Hole, B. Carr, C. W. Redman, A. L. Harris, and P. J. Dobson, “Sizing and phenotyping of cellular vesicles using Nanoparticle Tracking Analysis,” Nanomedicine 7(6), 780–788 (2011). [CrossRef]
15. W. Oosthuyzen, N. E. Sime, J. R. Ivy, E. J. Turtle, J. M. Street, J. Pound, L. E. Bath, D. J. Webb, C. D. Gregory, and M. A. Bailey, “Quantification of human urinary exosomes by nanoparticle tracking analysis,” J Physiol 591(23), 5833–5842 (2013). [CrossRef]
16. C. Moore, R. Wing, T. Pham, and J. V. Jokerst, “Multispectral nanoparticle tracking analysis for the real-time and label-free characterization of amyloid-β self-assembly in vitro,” Anal. Chem. 92(17), 11590–11599 (2020). [CrossRef]
17. Y. Dai, S. Bai, C. Hu, K. Chu, B. Shen, and Z. J. Smith, “Combined morpho-chemical profiling of individual extracellular vesicles and functional nanoparticles without labels,” Anal. Chem. 92(7), 5585–5594 (2020). [CrossRef]
18. C. F. Bohren and D. R. Huffman, Absorption and Scattering of Light by Small Particles (John Wiley & Sons, 2008).
19. X. Su, C. Capjack, W. Rozmus, and C. Backhouse, “2D light scattering patterns of mitochondria in single cells,” Opt. Express 15(17), 10562–10575 (2007). [CrossRef]
20. X. Su, S. Liu, X. Qiao, Y. Yang, K. Song, and B. Kong, “Pattern recognition cytometry for label-free cell classification by 2D light scattering measurements,” Opt. Express 23(21), 27558–27565 (2015). [CrossRef]
21. L. Zhang, Y. Xie, Y. Tu, L. Luo, K. Li, L. Yuan, W. Chen, H. Zhao, and Z. Zhang, “Clinical lymphocytes construction for light scattering inversion study: a three-dimensional morphology constructed method from defective confocal images,” J. Biomed. Opt. 23(08), 085003 (2018). [CrossRef]
22. J. Sun, L. Wang, Q. Liu, A. Tárnok, and X. Su, “Deep learning-based light scattering microfluidic cytometry for label-free acute lymphocytic leukemia classification,” Biomed. Opt. Express 11(11), 6674–6686 (2020). [CrossRef]
23. W. Y. Wan, L. Liu, X. Liu, W. Wang, M. Z. Islam, C. Dong, C. R. Garen, M. T. Woodside, M. Gupta, and M. Mandal, “Integration of light scattering with machine learning for label free cell detection,” Biomed. Opt. Express 12(6), 3512–3529 (2021). [CrossRef]
24. R. Parthasarathy, “Rapid, accurate particle tracking by calculation of radial symmetry centers,” Nat. Methods 9(7), 724–726 (2012). [CrossRef]
25. J. M. Newby, A. M. Schaefer, P. T. Lee, M. G. Forest, and S. K. Lai, “Convolutional neural networks automate detection for tracking of submicron-scale particles in 2D and 3D,” Proc. Natl. Acad. Sci. U. S. A. 115(36), 9026–9031 (2018). [CrossRef]
26. S. Helgadottir, A. Argun, and G. Volpe, “Digital video microscopy enhanced by deep learning,” Optica 6(4), 506–513 (2019). [CrossRef]
27. Z. Wang, S. Wang, G. Chen, and X. Su, “Deep learning based label-free small extracellular vesicles analyzer with light-sheet illumination differentiates normal and cancer liver cells,” Sens. Actuators, B 347, 130612 (2021). [CrossRef]
28. K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv, arXiv:1409.1556 (2014). [CrossRef]
29. A Krizhevsky, I Sutskever, and GE Hinton, “Imagenet classification with deep convolutional neural networks,” Commun. ACM 60(6), 84–90 (2017). [CrossRef]
30. T. Van Steenkiste, J. Ruyssinck, O. Janssens, B. Vandersmissen, F. Vandecasteele, P. Devolder, E. Achten, S. Van Hoecke, D. Deschrijver, and T. Dhaene, “Automated assessment of bone age using deep learning and Gaussian process regression,” in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), (IEEE, 2018), pp. 674–677.
31. B. H. van der Velden, M. H. Janse, M. A. Ragusi, C. E. Loo, and K. G. Gilhuijs, “Volumetric breast density estimation on MRI using explainable deep learning regression,” Sci. Rep. 10(1), 1–9 (2020). [CrossRef]
32. S. Wieser and G. J. Schütz, “Tracking single molecules in the live cell plasma membrane: Do’s and Don’t’s,” Methods 46(2), 131–140 (2008). [CrossRef]
33. D. Ernst and J. Köhler, “How the number of fitting points for the slope of the mean-square displacement influences the experimentally determined particle size distribution from single-particle tracking,” Phys. Chem. Chem. Phys. 15(10), 3429–3432 (2013). [CrossRef]
34. T. Wagner, H.-G. Lipinski, and M. Wiemann, “Dark field nanoparticle tracking analysis for size characterization of plasmonic and non-plasmonic particles,” J. Nanopart. Res. 16(5), 1–10 (2014). [CrossRef]
35. H. Saveyn, B. De Baets, O. Thas, P. Hole, J. Smith, and P. Van Der Meeren, “Accurate particle size distribution determination by nanoparticle tracking analysis based on 2-D Brownian dynamics simulation,” J. Colloid Interface Sci. 352(2), 593–600 (2010). [CrossRef]
36. A. Kim, W. Bernt, and N.-J. Cho, “Improved size determination by nanoparticle tracking analysis: influence of recognition radius,” Anal. Chem. 91(15), 9508–9515 (2019). [CrossRef]
37. R. Molenaar, S. Chatterjee, B. Kamphuis, I. M. Segers-Nolten, M. M. Claessens, and C. Blum, “Nanoplastic sizes and numbers: quantification by single particle tracking,” Environmental Science: Nano 8, 723 (2021). [CrossRef]
38. J.G. Walker, “Improved nano-particle tracking analysis,” Meas. Sci. Technol. 23(6), 065605 (2012). [CrossRef]
39. D. Ershov, M.-S. Phan, J. W. Pylvänäinen, S. U. Rigaud, L. Le Blanc, A. Charles-Orszag, J. R. Conway, R. F. Laine, N. H. Roy, and D. Bonazzi, “TrackMate 7: integrating state-of-the-art segmentation algorithms into tracking pipelines,” Nat. Methods 19(7), 1–4 (2022). [CrossRef]
40. J. H. Ryu, S. Y. Bang, J.-W. Yoon, C. S. Lim, and K. B. Shim, “Pulsed laser induced synthesis of scheelite-type colloidal nanoparticles in liquid and the size distribution by nanoparticle tracking analysis,” Appl. Surf. Sci. 253(20), 8408–8414 (2007). [CrossRef]
41. P. V. D. Meeren, M. Kasinos, and H. Saveyn, “Relevance of two-dimensional Brownian motion dynamics in applying nanoparticle tracking analysis,” in Nanoparticles in Biology and Medicine (Springer, 2012), pp. 525–534.
