Open Access
Add to BibSonomy
Abstract
Microalgae have recently been gaining attention for their versatile uses and environmentally friendly benefits. Accurate characterization and classification of a large population of microalgal cells with single-cell resolution are highly valuable for their diverse applications such as water treatment, biofuel production, food, and nitrogen-fixing biofertilization. Here we demonstrate accurate classification of spherical microalgal species using recently developed frequency-division-multiplexed fluorescence imaging flow cytometry and machine learning. We obtained three-color (bright-field and two-color fluorescence) images of microalgal cells, quantified morphological features of the cells using the images, and classified six microalgae using features via a support vector machine. By virtue of the rich information content of the three-color images of microalgal cells, we classified six microalgae with a high accuracy of 99.8%. Our method can evaluate large populations of microalgal cells with single-cell resolution and hence holds promise for various applications such as environmental monitoring of the hydrosphere.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Microalgae have recently been gaining attention because of their versatile uses and environmentally friendly benefits. Microalgae are unicellular species, which have a rather large range of size and biodiversity. They are incredibly important organisms as they produce oxygen and absorb harmful chemicals. Furthermore, they are the base of the food web for many marine-based ecosystems [1,2]. Additionally, they have diverse applications including water treatment, biofuel production, food, and nitrogen-fixing biofertilization [3–6]. In these applications, accurate characterization and classification of a large population of microalgal cells with single-cell resolution are highly valuable. For example, separating out certain strains of microalgae that are suitable for biofuel production could drastically increase the efficiency as a fuel source while keeping the price competitive with that of fossil fuels [6].
In a practical situation, such as monitoring bodies of water for ecological studies, classifying microalgae populations can be difficult due to the various types of microalgae, and similarities of morphology [7]. Determining these microalgae population variations is crucial for understanding how energy is transferred from the base of the aquatic food chain [8]. Furthermore, monitoring the microalgae populations in natural bodies of water is an imperative process for determining changes in ecosystems [9,10]. Separating the microalgae before classifications increases the required time, and would still require a validation step with a microscope. The morphological similarities and biodiversity of microalgae increase the difficulty of classification which translates to an increased analysis time. Recent methodologies, such as imaging flow cytometry, can reduce the required analysis time [11]. In particular, time-stretch optofluidic microscopy [12–16] has been shown to be effective for this purpose [15], but the lack of molecular-specific imaging (e.g., fluorescence imaging) and the consequent insufficiency of classification accuracy limit its utility.
In this study, we demonstrate an accurate classification of microalgal cells using recently developed frequency-division-multiplexed (FDM) fluorescence imaging flow cytometry [17–20] and machine learning. This system greatly enhances the throughput of fluorescence imaging flow cytometry by acquiring images with a single-pixel photodetector instead of a CCD camera. Using morphological parameters extracted from three-color (bright-field and two-color fluorescence) images, we classified six spherical microalgal species using a support vector machine (SVM), which is a typical machine learning method. By virtue of the three-color imaging capability, we achieved a high classification accuracy of 99.8%, which exceeds the classification accuracy values of conventional methods such as bright-field imaging flow cytometry (89.5%) and non-imaging fluorescence flow cytometry (84.9%).
The many morphological parameters acquired with the fluorescence imaging flow cytometer and the automation of the analysis overcome the difficulty with similar featured microalgae that previously limited the classification of microalgae. The utility of the method can be extended further by combining it with a cell sorter for intelligent image-activated cell sorting [19,20]. Moreover, this method holds promise for simultaneous evaluation of multiple microalgal cells while maintaining the capabilities of single-cell analysis [21].
2. Materials and methods
As shown in Figure 1(a), our procedure for high-speed accurate classification of microalgae by intelligent FDM fluorescence imaging flow cytometry is as follows. A sample of microalgae is pre-processed (fluorescence staining and concentration adjustment) and injected into a microfluidic chip so that multi-color images of the cells can be recorded at a flow speed of up to 2 m/s, which is controlled by a syringe pump. A 3D hydrodynamic focusing channel is employed to align the cells and avoid out-of-focus images. This imaging process is repeated for each of the microalgal species. Then, the image data for each species is combined into a single image dataset. The obtained image dataset is digitally processed by a computer and analyzed by a machine learning method for accurate classification using the rich information content of the multi-color images.
Fig. 1. High-speed accurate classification of microalgae by intelligent FDM fluorescence imaging flow cytometry. (a) Procedure. (b) Schematic of the FDM fluorescence imaging flow cytometer. HWP, half-wave plate; PBS, polarizing beam splitter; AOD, acousto-optic deflector; HBS, half beam splitter; DM, dichroic mirror; APD, avalanche photodetector; OL, objective lens; ND, neutral density filter. The inset shows an enlarged schematic of the flow channel and excitation beam spots inside the channel. (c) Flow chart of digital image processing.
The imaging part of the FDM fluorescence imaging flow cytometer is schematically shown in Figure 1(b) [17]. The output beam from a laser source (Coherent Genesis CX 488 STM, λ = 488 nm) is incident on a Mach-Zehnder interferometer that includes two acousto-optic deflectors (AODs) (Brimrose TED-300-200-488) driven by multi-tone signals (200-300 MHz and 300-400 MHz). An excitation beam array is created by the interference between the beam arrays deflected by the AODs. This excitation beam array forms a linear array of ∼50 beam spots with a 42-µm width, each of which is intensity-modulated at a different frequency. The beam array is focused on a microfluidic chip through an objective lens (OL) (Olympus UPLSAPO20X, NA = 0.75). The fluorescence light from each cell passing through the beam spot array is split into shorter (green) and longer (red) wavelength components by a dichroic mirror (DM) (Semrock ff580-fdi01, cutoff wavelength: 580 nm), each of which is detected by an avalanche photodetector (APD) (Thorlabs APD430A2/M). Similarly, the transmitted excitation beam array is detected by another APD (Thorlabs APD430A/M). The light intensity signals from the APDs are recorded by a digitizer (Spectrum M4i.2212-x8, 1.25 GSamples/s). Three-color images (two-color fluorescence and bright-field) of cells are created from the recorded signals via our proprietary digital lock-in detection algorithm implemented in LabVIEW.
We used six microalgal species from the Microbial Culture Collection at the National Institute for Environmental Studies [22] and Tsuruoka Keio Algae Collection of T. Nakada at the Institute for Advanced Biosciences at Keio University for our proof-of-concept demonstration: Scenedesmus aff. acutus (TKAC-1007), Gloeomonas anomalipyrenoide (NIES-3640), Chlamydomonas reinhardtii (TKAC-1017), Hamakko caudatus (NIES-2293), Chlorella sorokiniana (TKAC-1027), and Haematococcus lacustris (NIES-4141). Microalgal species were separately grown in culture flasks (working volume: 20 mL) in a 14h/10h light/dark cycle with illumination at 120-140 µmol photons/m2/s at 25°C using AF-6 medium [23]. Before flow imaging, cells were stained with 3 µM of SYTO 16 (Thermo Fisher Scientific, nucleus stain) in the culture medium followed by incubation in dark for 45 minutes.
As shown in Figure 1(c), our digital image processing flow is described as follows. First, raw images were pre-processed, which consists of low-pass filtering to reduce noise, truncating vacant areas to reduce data size, and creating tiled images to simplify the subsequent process. Then, morphological features of each image channel were calculated after the object areas were segmented. The segmentation was based on a mean correlation threshold. 251 morphological feature values from each cell image were extracted. These processes were performed by CellProfiler, an open source software for cellular image analysis [24]. Finally, we classified different types of cells via a sequential minimal optimization SVM [25,26] with a polynomial kernel using the feature values and WEKA, a free software program for machine learning [27]. The performance of the SVM was measured with a 10-fold cross-validation. The accuracy of the classification was determined by comparing the predicted results of the SVM analysis with the known ground truth. In addition, the weight of parameters for the classification result of the SVM were obtained, which shows which parameters mainly contribute to the classification.
3. Results
Images of the six microalgal species obtained by the FDM imaging flow cytometer are shown in Figure 2, where the green, red, and gray pseudocolors correspond to fluorescence of the nucleus stained by SYTO16, autofluorescence of the chlorophyll, and light diffraction (bright-field), respectively. Together, they elucidate the morphological structures of the chlorophyll and nucleus in each cell. By virtue of the high-speed imaging capability of the FDM fluorescence imaging flow cytometer, the images were blur-free even at a high-speed flow of 2 m/s. The image data acquisition of all channels was triggered when the raw signal level of the nucleus fluorescence channel exceeded that of the background level. A total of 8,640 cells were imaged (1,440 per species) with three images per cell (one for each channel), obtaining 25,920 images used for the analysis.
Fig. 2. Three-color images of the six microalgal species obtained by the FDM fluorescence imaging flow cytometer. Green represents the nucleus stained with SYTO 16. Red represents autofluorescent chlorophyll. Gray represents bright-field images. (a) Chlorella sorokiniana, (b) Chlamydomonas reinhardtii, (c) Haematococcus lacustris, (d) Hamakko caudatus, (e) Scenedesmus aff. acutus, (f) Gloeomonas anomalipyrenoide. The arrows indicate the flow direction. Color scales have been adjusted per species. Scale bars: 10 µm.
The classification results based on the morphological features obtained by the SVM are shown in Figure 3. With all of the 251 morphological features obtained by the three-color images, we obtained a classification accuracy of 99.8% [Figure 3(a)]. On the other hand, with only 87 of the morphological features extracted from the bright-field images, which simulated previously reported imaging flow cytometry that uses either time-stretch microscopy [28] or a CMOS image sensor without fluorescence imaging (i.e., bright-field imaging flow cytometry) [29,30], the classification accuracy dropped to 89.5% [Figure 3(b)]. Moreover, in the case where integrated pixel intensities of the three image channels were used so that no morphological information was given to the SVM, which corresponds to conventional non-imaging flow cytometry, the classification accuracy was found to be 84.9% [Figure 3(c)]. These results indicate that the multi-color (bright-field and fluorescence) imaging capability of the FDM fluorescence imaging flow cytometer plays a critical role in making the classification accuracy higher than the conventional methods such as bright-field imaging flow cytometry and non-imaging flow cytometry.
Fig. 3. Classification results corresponding to the number of cells that were classified for each species. (a) Confusion matrix showing the results of all 251 parameters with an accuracy of 99.8%. (b) Confusion matrix results showing the classification of the bright-field imaging flow cytometry results having an accuracy of 89.5%. (c) Confusion matrix results showing the classification of the non-imaging flow cytometry results having an accuracy of 84.9%.
4. Discussion
Inspecting the morphological features that make major contributions to the aforementioned high accuracy of the classification provides insights to the classification. Table 1 shows the top 10 features and their weights on the classification within the 251 features (see Tables 2–4 in the Appendix). A weight value, which is determined by calculating the Pearson’s correlation between the class and the features [31], quantitatively evaluates the impact of a feature on the classification compared with lower weight values. Most of the heavily weighted features were purely morphological ones without information regarding intensity, indicating the significance of the morphological information in classification. More specifically, morphological information of chlorophyll contributed the most within the three image channels. On the other hand, the integrated intensity edge (the integration of the pixel intensity along the perimeter of the nucleus) of the nucleus was the only top weighted parameter that was not morphologically dependent. This suggests that the intensity variation of the nuclear fluorescence was more important in the classification than its morphological counterparts, presumably due to the difference in staining affinity for the microalgal species.
Table 1. Top 10 features of the classification.
Low-dimensional plots such as a scatter plot or a histogram help to effectively elucidate and comprehend the differences and similarities between the microalgal species. For example, a scatter plot with the maximum Feret diameter of bright-field images and the mean intensity edge (the mean of the pixel intensity along the perimeter) of nucleus images is shown in Figure 4(a). While some of the species were clearly separated in the plot, there were large overlaps in two of the species (C. reinhardtii and C. sorokiniana). However, these species were clearly discriminated using the mean intensity of chlorophyll as shown in Figure 4(b). More information regarding the size of the nucleus and staining affinity can be seen from the scatter plot. For example, H. lacustris and G. anomalipyrenoide have similar Feret diameters while having vastly different nuclear intensities. The results suggest that our method can evaluate large populations of microalgal cells with single-cell resolution, and due to the high throughput nature of this technique, can further elucidate biological diversity in large populations of microalgae where higher throughput is necessary.
Fig. 4. Statistical plots of the morphological features of the six microalgal species. (a) 2D scatter plot with two parameters (Feret diameter and nucleus mean intensity edge). (b) Histogram of the congested region in the 2D scatter plot, showing how an additional parameter assists the classification.
Recently, a convolutional neural network (CNN) has been employed to classify cell types as a straightforward image-based classification methodology [19], but the present method of classification offers an advantage over a CNN in certain situations. Unlike the CNN-based method, our SVM based classification employs quantified morphological features that enables studies on the morphology of microalgae populations. Additionally, the morphological features obtained in the present method can provide insights into cellular heterogeneity of microalgal populations while providing biologically relevant information about the microalgal populations. On the other hand, the CNN-based method may be preferred if the classification accuracy is more important than rationales of the classification or further investigation of cellular heterogeneity within a single cell type.
Appendix
Table 2. Morphological features of microalgal cells in bright-field images (127 features).
Table 3. Top 10 features of the classification.
Table 4. Top 10 features of the classification.
Funding
Council for Science, Technology and Innovation (ImPACT Program); Japan Society for the Promotion of Science (Core-to-Core Program); White Rock Foundation; Precise Measurement Technology Promotion Foundation; Konica Minolta Science and Technology Foundation; Japan Society for the Promotion of Science (KAKENHI 19H05633).
Disclosures
H.M. and K.G. are inventors of a patent application covering FDM fluorescence imaging flow cytometry. K.G. is a shareholder of CYBO, Inc.
References
1. O. Pulz and W. Gross, “Valuable products from biotechnology of microalgae,” Appl. Microbiol. Biotechnol. 65(6), 635–648 (2004). [CrossRef]
2. D. R. Georgianna and S. P. Mayfield, “Exploiting diversity and synthetic biology for the production of algal biofuels,” Nature 488(7411), 329–335 (2012). [CrossRef]
3. A. Vaishampayan, R. P. Sinha, D. P. Hader, T. Dey, A. K. Gupta, U. Bhan, and A. L. Rao, “Cyanobacterial biofertilizers in rice agriculture,” Bot. Rev. 67(4), 453–516 (2001). [CrossRef]
4. P. He, S. Xu, H. Zhang, S. Wen, Y. Dai, S. Lin, and C. Yarish, “Bioremediation efficiency in the removal of dissolved inorganic nutrients by the red seaweed, Porphyra yezoensis, cultivated in the open sea,” Water Res. 42(4-5), 1281–1289 (2008). [CrossRef]
5. G. C. Dismukes, D. Carrieri, N. Bennette, G. M. Ananyev, and M. C. Posewitz, “Aquatic phototrophs: efficient alternatives to land-based crops for biofuels,” Curr. Opin. Biotechnol. 19(3), 235–240 (2008). [CrossRef]
6. M. Hannon, J. Gimpel, M. Tran, B. Rasala, and S. Mayfield, “Biofuels from algae: challenges and potential,” Biofuels 1(5), 763–784 (2010). [CrossRef]
7. K. Lee, M. Eisterhold, F. Rindi, S. Palanisami, and P. Nam, “Isolation and screening of microalgae from natural habitats in the Midwestern United States of America for biomass and biodiesel sources,” J. Nat. Sci. Biol. Med. 5(19), 333–339 (2014). [CrossRef]
8. L. Graham, J. Graham, and L. Wilcox, Algae (Pearson, 2009)
9. W. Li, X. Xu, M. Fujibayashi, Q. Niu, N. Tanaka, and O. Nishimura, “Response of microalgae to elevated CO2 and temperature: impact of climate change on freshwater ecosystems,” Environ. Sci. Pollut. Res. 23(19), 19847–19860 (2016). [CrossRef]
10. X. Irigoien, J. Huisman, and R. Harris, “Global biodiversity patterns of marine phytoplankton and zooplankton,” Nature 429(6994), 863–867 (2004). [CrossRef]
11. M. Hildebrand, A. Davis, R. Abbriano, H. Pugsley, J. Traller, S. Smith, R. P. Shrestha, O. Cook, E. Sanchez-Alvarez, K. Manandhar-Shrestha, and B. Alderette, “Applications of imaging flow cytometry for microalgae,” Methods in Mol. Bio. 1389, 47–67 (2016). [CrossRef]
12. C. Lei, H. Kobayashi, Y. Wu, M. Li, A. Isozaki, A. Yasumoto, H. Mikami, T. Ito, N. Nitta, T. Sugimura, M. Yamada, Y. Yatomi, D. Di Carlo, Y. Ozeki, and K. Goda, “High-throughput imaging flow cytometry by optofluidic time-stretch microscopy,” Nat. Protoc. 13(7), 1603–1631 (2018). [CrossRef]
13. K. Goda, A. Ayazi, D. R. Gossett, J. Sadasivam, C. K. Lonappan, E. Sollier, A. M. Fard, S. C. Hur, J. Adam, C. Murray, C. Wang, N. Brackbill, D. Di Carlo, and B. Jalali, “High-throughput single-microparticle imaging flow analyzer,” Proc. Natl. Acad. Sci. U. S. A. 109(29), 11630–11635 (2012). [CrossRef]
14. K. Goda and B. Jalali, “Dispersive fourier transformation for fast continuous single-shot measurements,” Nat. Photonics 7(2), 102–112 (2013). [CrossRef]
15. Q. T. K. Lai, K. C. M. Lee, A. H. L. Tang, K. K. Y. Wong, H. K. H. So, and K. K. Tsia, “High-throughput time-stretch imaging flow cytometry for multi-class classification of phytoplankton,” Opt. Express 24(25), 28170–28184 (2016). [CrossRef]
16. K. Goda, K. K. Tsia, and B. Jalali, “Serial time-encoded amplified imaging for real-time observation of fast dynamic phenomena,” Nature 458(7242), 1145–1149 (2009). [CrossRef]
17. H. Mikami, J. Harmon, H. Kobayashi, S. Hamad, Y. Wang, O. Iwata, K. Suzuki, T. Ito, Y. Aisaka, N. Kutsuna, K. Nagasawa, H. Watarai, Y. Ozeki, and K. Goda, “Ultrafast confocal fluorescence microscopy beyond the fluorescence lifetime limit,” Optica 5(2), 117–126 (2018). [CrossRef]
18. E. D. Diebold, B. W. Buckley, D. R. Gossett, and B. Jalali, “Digitally synthesized beat frequency multiplexing for sub-millisecond fluorescence microscopy,” Nat. Photonics 7(10), 806–810 (2013). [CrossRef]
19. N. Nitta, T. Sugimura, A. Isozaki, H. Mikami, K. Hiraki, S. Sakuma, T. Iino, F. Arai, T. Endo, Y. Fujiwaki, H. Fukuzawa, M. Hase, T. Hayakawa, K. Hiramatsu, Y. Hoshino, M. Inaba, T. Ito, H. Karakawa, Y. Kasai, K. Koizumi, S. Lee, C. Lei, M. Li, T. Maeno, S. Matsusaka, D. Murakami, A. Nakagawa, Y. Oguchi, M. Oikawa, T. Ota, K. Shiba, H. Shintaku, Y. Shirasaki, K. Suga, Y. Suzuki, N. Suzuki, Y. Tanaka, H. Tezuka, C. Toyokawa, Y. Yalikun, M. Yamada, M. Yamagishi, T. Yamano, A. Yasumoto, Y. Yatomi, M. Yazawa, D. Di Carlo, Y. Hosokawa, S. Uemura, Y. Ozeki, and K. Goda, “Intelligent image-activated cell sorting,” Cell 175(1), 266–276.e13 (2018). [CrossRef]
20. A. Isozaki, H. Mikami, K. Hiramatsu, S. Sakuma, Y. Kasai, T. Iino, T. Yamano, A. Yasumoto, Y. Oguchi, N. Suzuki, Y. Shirasaki, T. Endo, T. Ito, K. Hiraki, M. Yamada, M. Matsusaka, T. Hayakawa, H. Fukuzawa, Y. Yatomi, F. Arai, D. Di Carlo, A. Nakagawa, Y. Hoshino, Y. Hosokawa, S. Uemura, T. Sugimura, Y. Ozeki, N. Nitta, and K. Goda, “A practical guide to intelligent image-activated cell sorting,” Nat. Protoc. 14(8), 2370–2415 (2019). [CrossRef]
21. H. Mikami, C. Lei, N. Nitta, T. Sugimura, T. Ito, Y. Ozeki, and K. Goda, “High-speed imaging meets single-cell analysis,” Chem 4(10), 2278–2300 (2018). [CrossRef]
22. M. M. Watanabe, M. Kawachi, M. Hiroki, and F. Kasai, “NIES Collection List of Strains, 6th ed.” (NIES, 2000)
23. F. Kasai, M. Kawachi, M. Erata, F. Mori, K. Yumoto, M. Sato, and M. Ishimoto, “NIES-Collection List of Strains, 8th ed.,” Jpn. J. Phycol. 57(Suppl), 1–350 (2009).
24. T. Blasi, H. Hennig, H. D. Summers, F. J. Theis, J. Cerveira, J. O. Patterson, D. Davies, A. Filby, A. E. Carpenter, and P. Rees, “Label-free cell cycle analysis for high-throughput imaging flow cytometry,” Nat. Commun. 7(1), 10256 (2016). [CrossRef]
25. V Keeman, “Support Vector Machines – An Introduction,” in Support Vector Machines: Theory and Applications, Wang L. (Springer, 2005)
26. N. Cristianini and J. Shawe-Taylor, “An introduction to support vector machines: and other kernel-based learning methods,” (Cambridge University, 2000), Chap. 7.5.
27. E. Frank, M. A. Hall, and I. H. Witten, “Data mining, fourth edition: practical machine learning tools and techniques,” (Morgan Kaufmann, 2016)
28. C. Lei, T. Ito, M. Ugawa, T. Nozawa, O. Iwata, M. Maki, G. Okada, H. Kobayashi, X. Sun, P. Tiamsak, N. Tsumura, K. Suzuki, D. Di Carlo, Y. Ozeki, and K. Goda, “High-throughput label-free image cytometry and image-based classification of live Euglena gracilis,” Biomed. Opt. Express 7(7), 2703–2708 (2016). [CrossRef]
29. F. Kubota, H. Kusuzawa, T. Kosaka, and H. Nakamoto, “Flow cytometer and imaging device used in combination,” Cytometry 21, 129–132 (1995). [CrossRef]
30. F. Kubota, “Analysis of red cell and platelet morphology using an imaging-combined flow cytometer,” Clin. Lab. Haematol. 25(2), 71–76 (2003). [CrossRef]
31. L. Lin, “A concordance correlation coefficient to evaluate reproducibility,” Biometrics 45(1), 255–268 (1989). [CrossRef]
