We develop a pattern recognition cytometric technique for label-free cell classification. Two dimensional (2D) light scattering patterns from single cells and cell aggregates are obtained with a static cytometer. Good performance of the cytometric setup is verified by comparing yeast cell experimental results with theoretical simulations. Adaptive boosting (AdaBoost) method (a machine learning algorithm) is adopted for the analysis of the 2D light scattering patterns. It is shown that aggregates of three yeast cells can be well differentiated from aggregates of four yeast cells by this pattern recognition cytometric technique. We demonstrate that the pattern recognition cytometry can perform label-free classification of normal cervical cells and HeLa cells with a high accuracy rate.
© 2015 Optical Society of America
Conventional flow cytometry measures optical signals from single biological cells or particles in a fluidic stream and obtains cell size, organelle, and molecular information . In these cytometers, multiple detectors such as photodiodes or photomultiplier tubes are used to measure the forward scattering or side scattering light signals. With the aid of specially developed fluorescence dyes, conventional flow cytometry can perform quantitative analysis on single cells with multiple parameters and a high throughput [1,2]. This has made the conventional flow cytometry widely used in biomedicine for cancer diagnostics, regenerative medicine, and immunology [1–4]. However the conventional flow cytometry lacks the capability for detailed study on single cells for that the obtained fluorescence or scattered light signals are the integrated light intensity from a single cell. In order to perform single cell analysis with high resolution, imaging cytometric technologies have been developed, which can provide a moderate throughput and image information of the cells [5,6].
Another trend of optical flow cytometry development is for label-free cytometry. The first generation of flow cytometry in early 1970s measured the scattered light for single cell analysis in a label-free manner [7,8]. Since Mourant et al. reported that the scattered light, especially the side scattered light contains organelle information in single cells , light scattering methods for label-free cell analysis have been widely studied [10–13]. These reports focus on the one dimensional light scattering (the scattered light distribution in polar angle), but the two dimensional (2D) light scattering (the scattered light distribution in both polar and azimuthal angles) contains much richer information of the cells. Our group has developed the 2D light scattering microfluidic cytometry for particle and cell analysis [14,15], where we obtained both the 2D light scattering patterns via a lens-free system and the defocused scatter images through a microscope objective. And the latter was further developed as a microscope-based label-free microfluidic cytometric technique . Recently, Qu et al. have developed the two-photon excited autofluorescence label-free cytometry, which measures the strong autofluorescence emission from plasma for blood vessel visualization and blood cell counting [17,18].
The obtained results from 2D light scattering cytometry have been analyzed by Fourier transform methods for particle or cell size determination [15,19]. For the analysis of cell organelles, we have reported that mitochondria may dominate the 2D light scattering patterns from single cells . Analysis of the blobs (small scale 2D structures in the patterns) has been demonstrated by us for the differentiation of cells [16,20]. Mourant et al. have also studied the correlation between cell organelles and the 2D light scattering patterns . Although our previous reports have shown that the 2D light scattering cytometry can be used for cell diagnostics, it has not yet been developed to classify unknown cells. Considering what obtained in the 2D light scattering cytometry are patterns, pattern recognition [22,23] may broaden the applications of the 2D light scattering cytometry for cell diagnostics, such as for the automatic classification of unknown cells with supervised studies (using the 'labeled' training data, that is, the patterns are from cells that were verified).
In this manuscript we develop a pattern recognition cytometric technique for label-free classification of cells. Adaptive boosting (AdaBoost) method [24,25], which has the potential for cell classification by using multiple weak learners, is adopted here to perform pattern recognition of the 2D light scattering results (not the cell images as obtained using microscope or imaging flow cytometer). The experimental patterns from cells are obtained using our recently developed 2D light scattering static cytometry . By using this static cytometer, the aggregates of yeast cells with known cell numbers can be studied, which is challenging to achieve with flow cytometry. The yeast cell aggregates are used to model the 2D light scattering from organelles in human cells. It has been demonstrated in this manuscript that aggregates of three yeast cells can be well differentiated from aggregates of four yeast cells. High accuracy rate for the label-free classification of normal cervical cells and HeLa cells has been obtained with the pattern recognition cytometry, which is promising for cervical cancer screening as compared with the commonly used methods [27–30].
2. Materials and methods
2.1 Materials and experimental setup
Figure 1 shows a schematic diagram for the experimental setup of our pattern recognition cytometer. As shown in Fig. 1, a linearly polarized laser beam from a diode pumped solid state (DPSS) laser (Frankfurt Laser Company, 100 mW, Germany) with an emission wavelength of 532 nm is coupled into one end of an optical fiber (Thorlabs, 105/125 μm, USA) via a 4 × microscope objective with a numerical aperture (NA) of 0.1. The other end of this fiber is mounted on a 3D translation stage, which is used for the illumination of single cells (as illustrated in Fig. 1) or cell aggregates on chip. The system for the illumination of cells using an optical fiber and the obtaining of 2D light scattering patterns with a CMOS sensor (Canon, APS-C, Japan) is facilitated with a microscope (Olympus, BX 53, Japan). The 2D light scattering patterns are obtained when the microscope is working in defocus mode [15,16,20]. If not otherwise specified, the microscope is defocused about 200 μm, and the integration time used for the CMOS sensor is 1/25s. For the experimental setup in Fig. 1, the obtained 2D light scattering patterns are in the polar and azimuthal angular ranges of approximately 22 degrees via a 10 × microscope objective with an NA of 0.25. Pattern recognition is then performed for the classification of cells with different aggregations, or for the classification of normal cervical cells from HeLa cells.
2.2 Cell sample preparation
The yeast cell solution was made by dissolving dry yeasts in ultrapure water and sonicated to get a homogeneous sample. The concentration of the yeast cell solution is around 3,000 cells/mL for single cell analysis.
Normal cervical cells are derived from the remaining cells after TCT (Thinprep Cytologic Test, Hologic, USA). The cell samples are without human papillomavirus (HPV) infection, chronic cervicitis and any other cervical lesions. After the normal cervical tissues were made into paste, 3 mL of 0.25% trypsin-EDTA (Sigma, USA) was added. The paste was left for 20 minutes (mins) at 37°C with 5% CO2, and stirred every 5 mins. Once a jelly-like sample was obtained, the supernatant was removed and 3.0 mL of 0.2% collagenase (Sigma, USA) was added. The sample was again left at 37°C and stirred for 30 mins. Then 3 mL DMEM (Dulbecco's Modified Eagle's Medium, Gibco, USA) with 10% heat-inactivated fetal bovine serum (Invitrogen, Gibco, USA) was added, and the sample was filtered with a 200-mesh strainer. The cell suspension was collected and centrifuged at 1,000 rpm for 5 mins (Thermo Primo, USA). After the supernatant was removed, 1mL 1x phosphate buffered saline (PBS) solution was added, and the cell suspension was centrifuged again. Finally the cells were re-suspended in 1mL 1x PBS solution after the supernatant was removed.
The HeLa cells were maintained in complete media consisting of RPMI-1640 (Roswell Park Memorial Institute, Hyclone, USA) with 10% heat-inactivated fetal bovine serum and 1% 10,000 IU/mL Penicillin and streptomycin at 37°C, 5% CO2, and 100% humidity. The HeLa cells were collected into a centrifuge tube, treated with 0.25% trypsin-EDTA (Sigma, USA), and centrifuged at 1,000 rpm for 5 mins. Once the supernatant was removed, the cells were washed with 1mL 1x PBS solution, re-suspended and centrifuged. Finally, the final supernatant was removed, and the cells were re-suspended in 1mL 1x PBS.
The same procedure was used for the fixation of normal cervical cells and HeLa cells. The 1 mL final suspension was added into 4 mL 95% ethyl alcohol solution (Tieta Reagents, China) while vortexed at low speed, and fixed for 12 hours at 4°C. After fixation, the cell solution was centrifuged at 1,000 rpm for 5 mins, and the supernatant was removed. The cells were again washed, centrifuged, and re-suspended in 5 mL 1x PBS. At last, the final supernatant was removed and the cells were re-suspended in 1mL 1x PBS.
The fixed normal cervical cell or HeLa cell solution was transferred to a centrifuge tube, stored on ice. For single cell analysis, the concentration is around 3,000 cells/mL for both the normal cervical cell and HeLa cell solutions.
2.3 Validation of the pattern recognition cytometer
In order to validate the performance of the pattern recognition cytometry, we have performed experimental measurements on single yeast cells, which are bio-safe and accessible to many labs. Simulated 2D light scattering patterns from single yeast cells are obtained with our Mie theory algorithm . Figure 2 shows the comparison between the yeast cell experimental light scattering patterns and our Mie theory simulated results. As shown in Figs. 2(a) and 2(b) are the microscope images for different yeast cells with diameters around 3.8 and 5.0 µm, respectively. Figure 2(c) is the 2D light scattering patterns of the single yeast cell shown in Fig. 2(a), and Fig. 2(d) is the 2D pattern of Fig. 2(b). For Mie theory simulations, the yeast cells are assumed to be spherical with a refractive index of 1.42, surrounded by water with refractive index of 1.334. The incident laser beam is linearly polarized with a wavelength of 532 nm. The Mie theory simulated 2D light scattering patterns from yeast cell models with a diameter of 3.8 and 5.0 µm are shown in Figs. 2(e) and 2(f), respectively.
It is noticed that the simulated 2D light scattering patterns are with three and four fringes in the angular range from 79 to 101 degrees in polar angle. Experimentally we have also observed yeast cell light scattering patterns with three and four fringes as shown in Figs. 2(c) and 2(d). Previously we have reported that the fringe numbers in the 2D light scattering patterns vary linearly with the size variations of spherical scatterers , as similar as shown here in Fig. 2. The experimental patterns agree well with the Mie theory simulated 2D light scattering patterns, thus validates the performance of the pattern recognition cytometer.
3. Results and discussion
3.1 Pattern recognition cytometry for the differentiation of cell aggregates
Our recently developed 2D light scattering static cytometry makes the study of light scattering from cell aggregates practical . In the static cytometer the cells are not flowing but static on chip, and a scanning probe is used to illuminate the cells. Thus detailed light scattering study of better controlled cell samples can be obtained. In this manuscript, the yeast cells are manually controlled to form aggregates of three or four yeast cells. The diameters of these yeast cell aggregates are comparable to typical human cells.
Figure 3 shows the 2D light scattering patterns from different yeast cell aggregates. Figures 3(a) and 3(b) are the yeast cell aggregates with three or four cells. Note the yeast cell aggregates are about 10 μm in diameter. The 2D light scattering patterns of the yeast cell aggregates in Figs. 3(a) and 3(b) are shown in Figs. 3(c) and 3(d), respectively. Without detailed study, the cell aggregates with three or four yeast cells contribute to different 2D light scattering patterns.
To perform automatic classification of the different yeast cell aggregates, a data set of 60 2D light scattering patterns from 60 different yeast cell aggregates is obtained using our 2D light scattering static cytometer as demonstrated in Fig. 1. There are 30 cases for different yeast cell aggregates with three cells, and 30 cases for four cells. The dimension of each 2D light scattering pattern is cropped to 250 by 250 pixels, and all the cropped patterns are centered and normalized.
The leave-one-out experiments are performed using AdaBoost method. That is, we train 59 patterns with known contributors (either three-yeast-cell or four-yeast-cell aggregates) to obtain a set of weak learners, and leave the 60th pattern to be tested. The correct number (CN) for the testing is recorded and the accuracy rate (AR) is calculated using the following equation: , where TN is the total number of 2D light scattering patterns. Tendency of AR while increasing the number of weak learners has been studied, which shows that a highest accuracy rate of 86.7% can be obtained when 3 weak learners are used. Table 1 shows our pattern recognition cytometry for the classification of the different yeast cell aggregates. It is noticed that the AR for three-yeast-cell aggregates is 93.3%, and that for four-yeast-cell aggregates is 80%. This is reasonable because the aggregates with four yeast cells have more complex 3D distributions, and some cases of the four-yeast-cell aggregates may contribute similarly as the three-yeast-cell aggregates to the 2D light scattering patterns obtained using the static cytometer.
3.2 Pattern recognition cytometry differentiates normal cervical cells from HeLa cells
Having shown that the pattern recognition cytometry can be used to differentiate cell aggregates with known distributions, we then apply this technique for the classification of normal cervical cells from HeLa cells. Figure 4 shows the 2D light scattering patterns obtained from a normal cervical cell and a HeLa cell. Figures 4(a) and 4(b) are the microscope cell images, while Figs. 4(c) and 4(d) are the 2D light scattering patterns of Figs. 4(a) and 4(b), respectively. For the data set we used here to perform pattern recognition using AdaBoost method, there are 54 light scattering patterns from normal cervical cells, and 38 patterns from HeLa cells. Similar process is performed on the light scattering patterns from these two kinds of cervical cells, as described above for the yeast cell aggregates.
The leave-one-out experiments have also been performed here, however there are now 91 2D light scattering patterns to train the weak learners, and the 92th pattern is left for testing. There are altogether 92 experiments performed. Table 2 shows the classification results for the normal cervical cells and HeLa cells. Our experiments have shown that a highest AR of 90.2% can be obtained when 7 weak learners are used to form the classifier. It is found that the pattern recognition cytometry offers an AR of 90.7% for the classification of normal cervical cells and an AR of 89.5% for the HeLa cells. Our results show that when the normal cervical cells are of similar shapes with the HeLa cells as shown in Fig. 4, the inner structures of these two kinds of cells contribute distinctively to the 2D light scattering patterns. An overall AR of 90.2% shown in this manuscript is very promising for cervical cancer screening in clinics.
In summary, we reported here a pattern recognition cytometric technique for label-free cell classification by measurements of the 2D light scattering from single cells or cell aggregates. The pattern recognition cytometry adopted an AdaBoost method for the analysis of the 2D light scattering patterns. The AdaBoost method performed leave-one-out tests with automatic machine learning, which solidified our pattern recognition cytometry for cell classification. In this report, it was shown that the recognition accuracy rate for the differentiation of yeast cell aggregates with three or four cells was 86.7%. Note that the diameters of yeast cell aggregates are comparable with human white blood cells. Thus our results showed that the pattern recognition cytometry not only can differentiate cell aggregates, but also can be used to study organelle variations in single cells. The pattern recognition cytometry was applied to differentiating normal cervical cells from HeLa cells, and an accuracy rate of 90.2% was obtained which is very promising in cervical cancer screening. In the next step we will apply pattern recognition cytometry for cervical cancer screening with clinical trials.
Xuantao Su thanks the financial support from the National Natural Science Foundation of China (Grant No. 81271615).
References and links
1. H. M. Shapiro, Practical Flow Cytometry (John Wiley & Sons, Inc., 2003).
2. L. A. Herzenberg, D. Parks, B. Sahaf, O. Perez, M. Roederer, and L. A. Herzenberg, “The history and future of the fluorescence activated cell sorter and flow cytometry: a view from Stanford,” Clin. Chem. 48(10), 1819–1827 (2002). [PubMed]
6. U. Erdbrügger, C. K. Rudy, M. E Etter, K. A. Dryden, M. Yeager, A. L. Klibanov, and J. Lannigan, “Imaging flow cytometry elucidates limitations of microparticle analysis by conventional flow cytometry,” Cytometry A 85(9), 756–770 (2014). [CrossRef] [PubMed]
8. G. C. Salzman, J. M. Crowell, C. A. Goad, K. M. Hansen, R. D. Hiebert, P. M. LaBauve, J. C. Martin, M. L. Ingram, and P. F. Mullaney, “A flow-system multiangle light-scattering instrument for cell characterization,” Clin. Chem. 21(9), 1297–1304 (1975). [PubMed]
9. J. R. Mourant, J. P. Freyer, A. H. Hielscher, A. A. Eick, D. Shen, and T. M. Johnson, “Mechanisms of light scattering from biological cells relevant to noninvasive optical-tissue diagnostics,” Appl. Opt. 37(16), 3586–3593 (1998). [CrossRef] [PubMed]
10. V. Backman, R. Gurjar, K. Badizadegan, L. Itzkan, R. R. Dasari, L. T. Perelman, and M. S. Feld, “Polarized light scattering spectroscopy for quantitative measurement of epithelial cellular structures in situ,” IEEE J. Sel. Top. Quantum Electron. 5(4), 1019–1026 (1999). [CrossRef]
11. R. Drezek, A. Dunn, and R. Richards-Kortum, “Light scattering from cells: finite-difference time-domain simulations and goniometric measurements,” Appl. Opt. 38(16), 3651–3661 (1999). [CrossRef] [PubMed]
12. A. Wax, C. Yang, V. Backman, K. Badizadegan, C. W. Boone, R. R. Dasari, and M. S. Feld, “Cellular organization and substructure measured using angle-resolved low-coherence interferometry,” Biophys. J. 82(4), 2256–2264 (2002). [CrossRef] [PubMed]
13. X. Li, A. Taflove, and V. Backman, “Recent progress in exact and reduced-order modeling of light-scattering properties of complex structures,” IEEE J. Sel. Top. Quantum Electron. 11(4), 759–765 (2005). [CrossRef]
15. X. T. Su, K. Singh, C. Capjack, J. Petrácek, C. Backhouse, and W. Rozmus, “Measurements of light scattering in an integrated microfluidic waveguide cytometer,” J. Biomed. Opt. 13(2), 024024 (2008). [CrossRef] [PubMed]
16. X. Su, S. E. Kirkwood, M. Gupta, L. Marquez-Curtis, Y. Qiu, A. Janowska-Wieczorek, W. Rozmus, and Y. Y. Tsui, “Microscope-based label-free microfluidic cytometry,” Opt. Express 19(1), 387–398 (2011). [CrossRef] [PubMed]
17. Y. Zeng, J. Xu, D. Li, L. Li, Z. Wen, and J. Y. Qu, “Label-free in vivo flow cytometry in zebrafish using two-photon autofluorescence imaging,” Opt. Lett. 37(13), 2490–2492 (2012). [CrossRef] [PubMed]
18. Y. Zeng, B. Yan, Q. Sun, S. He, J. Jiang, Z. Wen, and J. Y. Qu, “In vivo micro-vascular imaging and flow cytometry in zebrafish using two-photon excited endogenous fluorescence,” Biomed. Opt. Express 5(3), 653–663 (2014). [CrossRef] [PubMed]
19. S. Yu, J. Zhang, M. S. Moran, J. Q. Lu, Y. Feng, and X.-H. Hu, “A novel method of diffraction imaging flow cytometry for sizing microspheres,” Opt. Express 20(20), 22245–22251 (2012). [CrossRef] [PubMed]
20. X. Su, Y. Qiu, L. Marquez-Curtis, M. Gupta, C. E. Capjack, W. Rozmus, A. Janowska-Wieczorek, and Y. Y. Tsui, “Label-free and noninvasive optical detection of the distribution of nanometer-size mitochondria in single cells,” J. Biomed. Opt. 16(6), 067003 (2011). [CrossRef] [PubMed]
22. A. K. Jain, R. P. W. Duin, and J. C. Mao, “Statistical pattern recognition: a review,” IEEE Trans. Pattern Anal. 22(1), 4–37 (2000). [CrossRef]
23. P. Viola and M. J. Jones, “Robust real-time face detection,” Int. J. Comput. Vis. 57(2), 137–154 (2004). [CrossRef]
24. Y. Freund, “Boosting a weak learning algorithm by majority,” Inf. Comput. 121(2), 256–285 (1995). [CrossRef]
25. Y. Freund and R. E. Schapire, “A decision-theoretic generalization of on-line learning and an application to boosting,” J. Comput. Syst. Sci. 55(1), 119–139 (1997). [CrossRef]
26. L. Xie, Y. Yang, X. Sun, X. Qiao, Q. Liu, K. Song, B. Kong, and X. Su, “2D light scattering static cytometry for label-free single cell analysis with submicron resolution,” Cytometry A (to be published). [PubMed]
27. D. Saslow, C. D. Runowicz, D. Solomon, A. B. Moscicki, R. A. Smith, H. J. Eyre, C. Cohen, and American Cancer Society, “American Cancer Society guideline for the early detection of cervical neoplasia and cancer,” CA Cancer J. Clin. 52(6), 342–362 (2002). [CrossRef] [PubMed]
28. J. Swan, N. Breen, R. J. Coates, B. K. Rimer, and N. C. Lee, “Progress in cancer screening practices in the United States: Results from the 2000 National Health Interview Survey,” Cancer 97(6), 1528–1540 (2003). [CrossRef] [PubMed]
29. R. Drezek, M. Guillaud, T. Collier, I. Boiko, A. Malpica, C. Macaulay, M. Follen, and R. Richards-Kortum, “Light scattering from cervical cells throughout neoplastic progression: influence of nuclear morphology, DNA content, and chromatin texture,” J. Biomed. Opt. 8(1), 7–16 (2003). [CrossRef] [PubMed]