1. INTRODUCTION

A central challenge of biology is to correlate the phenotype of heterogeneous individuals in a population to their genotype in order to understand whether they conform to the observed population behavior or stand out as exceptions that drive disease or become threats to health [1–4]. While optical microscopy is a cornerstone method to study the morphology and molecular composition of biological specimens, flow cytometry is a gold standard for quantitative high-throughput single-cell characterization in numerous biomedical applications [5,6]. Recognizing the need to merge these two powerful platforms, several groups have proposed techniques for imaging flow cytometry (IFC) [7]. IFC simultaneously produces ensemble-averaged measurements and high-content spatial metrics from individual cells in a large population of cells without perturbation due to experiment condition change. A significant limitation of existing IFC systems is that, regardless of the optical detection method and computation algorithm is used, only two-dimensional (2D) cell images can be obtained [8–10]. The absence of three-dimensional (3D) tomography results in occlusion of objects, blurring by focal depth, loss of $z$-axis spatial resolution, and artifacts due to the projection of a 3D cell into a 2D image. For example, with 2D microscopic imaging, if a fluorescent probe is observed at the center of a cell, its location (e.g., membrane, cytosol, nucleus) is ambiguous. For a range of applications, such as internalization measurements, probe co-localization, and spot counting, relative to 2D imaging that is dependent on the cellular orientation to the imaging plane, 3D images provide more complete and accurate phenotyping of cell and organelle morphology, as well as nucleic acid and protein localization to support biological insights [11]. However, rapid and continuous 3D image acquisition for single cells in the microfluidic environment has not been achieved.

Recent research attempted to utilize 3D microscopy technology with microfluid compactible configuration to achieve 3D cell tomography in a microfluidic environment. Wu et al. reports a light-sheet fluorescence microscope with a microfluid-compatible arrangement at a cell flow rate of 1 mm/s and an image acquisition time of 22.5 ms per cell [12]. Quint et al. demonstrates a 3D confocal microscopy approach with a tilted microfluidic arrangement to capture cross-sectioning images with a maximum flow rate of 1.5 mm/s [13]. Sung et al. and Merola et al. present tomographic phase microscopy approaches for 3D refractive index distribution measurement using a CCD camera. However, it takes seconds for image recovery because of the computation complexity [14,15]. Although the above attempts achieved reasonable image resolution, the imaging throughput falls far below the applicable range. Recently, Martin et al. demonstrated a line excitation array detection (LEAD) fluorescence microscopy to implement high-throughput cross-sectioning imaging at 1 m/s flow rate with 0.8 million frames per second through leveraging fast laser scanning and photomultiplier detector array [16]. Although the LEAD fluorescence microscopy achieves a spatial resolution of 3.5 μm, the image resolution is limited by both the excitation beam size and the detector array size. In addition, LEAD fluorescence microscopy does not perform label-free detection.

Here we demonstrate high-throughput 3D imaging flow cytometry (3D-IFC) based on optical sectioning microscopy [17]. This combination of light-sheet scanning illumination technique and spatial-temporal transformation detection technique enables fluorescent and label-free 3D cell image reconstruction from single-element photodetector readout without a camera [18,19]. Building upon the speed and sensitivity benefits of the photomultiplier tube (PMT), the 3D-IFC uses multiple scanning techniques to add spatial information in a conventional flow cytometry architecture. 3D imaging is achieved by laser scanning across the first ($z$) axis, the cell translating by flow across the second ($y$) axis, and the use of multiple pinholes arranged along the third ($x$) axis to produce fluorescent and label-free information from 6000 voxels per scanning volume. By precisely mapping time to space, photodetector readout at one time point corresponds to one voxel in a 3D space. Here we demonstrate 3D-IFC of fluorescence and 90 deg label-free side-scattering (SSC) imaging of single cells in the microfluidic channel at a velocity of 0.2 m/s, corresponding to a throughput of approximately 500 cells per second.

2. PRINCIPLE AND METHODS

A schematic of the 3D-IFC system is shown in Fig. 1(a). In the 3D-IFC system, suspended cells form a 2D hydrodynamically focused single file in a quartz flow cell with a square cross section (Fig. 1) [20]. Laser excitation is via a light sheet ($x-y$ plane) with a diffraction-limited beam waist and a height of 200–400 μm, scanning in the $z$ direction at 200 kHz. When a cell is flowing through the whole optical interrogation at 0.2 m/s, a pixelated field of view is represented by a 3D space with $X$ by $Y$ by $Z$ voxels, as shown in Fig. 1(c). A pinhole array on the spatial filter is aligned at a tilting angle, $\vartheta$, to the flow stream, so the pinhole array also steps along the $x$ direction. In this manner, each pinhole allows light from voxels with a distinct $x$ index to reach the PMT detector (see Fig. 2 for details of the pinhole mask design). The imaging process begins when a flowing cell appears at the first pinhole of the spatial filter. During the first light-sheet scanning period (5 μs), the light intensity of voxels $z_{1-Z}$ with $x_1{y}_1$ index is collected. As the cell flows downstream in the $y$ direction to the next position, $x_1{y}_2$, the corresponding $z_{1-Z}$ voxels are produced. In this manner, when a cell completely passes pinhole 1, the entire 2D $yz$ slice at $x_1$ is imaged. As the cell travels farther downstream the $y$ direction and passes through all the following pinholes, $yz$ slices of at $x_2$ to $x_X$ are recorded.

 

Fig. 1. Implementation of the 3D-IFC System and demonstration of time to 3D space mapping. (a) Schematic diagram of the 3D-IFC system. AOD, acousto-optic deflector; CL, cylindrical lens; IO, 50X/0.55 illumination objective lens; DO, 10X/0.28 detection objective lens; SF, spatial filter; DMs, dichroic mirrors; PMT, photomultiplier tube; DIG, $125{\mathrm{MSs}}^{-1}$ digitizer. The AOD and CL produce a scanning light sheet. The sample is 2D hydrodynamically focused by sheath before entering the square cross section quartz flow cell. (b) Optical interrogation area. $H$, the height of the light sheet; $\vartheta$, tilt angle between flow ($y$ axis) and vertical line. Illumination light sheet propagates horizontally and scans in the $z$ axis; the sample flows in the $y$ axis; $x$ is the orthogonal axis. The spatial filter at the image plane uses pinholes to produce line scans across the $x$ axis. (c) 3D reconstructed space. The resolution on the $x$ axis is determined by the number of pinholes (pixelated field of view in the $x$ direction); resolution of $Y$ by the distance between two slits (pixelated field of view in the $y$ direction); and resolution of $Z$ by the light-sheet scanning range (pixelated field of view in the $z$ direction); (d) one light-sheet scan period produces a 1D light intensity profile in the $z$ axis. The PMT voltage readout of one sample point corresponds to the light intensity of one voxel in the $z$-axis. (e) While object travels along the $y$ axis, multiple scans produce a 2D profile in the $yz$ plane within one pinhole scan period. Each section—separated by dotted lines—corresponds to the light intensity of one row in the 2D image stack. (f) When an object completely passes through the spatial filter covering the area, the time-domain signal contains the complete information of the 3D profile in the $xyz$ space. Each section corresponds to one 2D image slice. AOD, tuning voltage of the AOD driver; FL1, PMT readout of fluorescence detection channel 1; FL2, PMT readout of fluorescence detection channel 2; SSC, PMT readout of SSC light detection channel.

Download Full Size | PDF

 

Fig. 2. Spatial filter design. Two examples of spatial filters placed at the image plane. The top two and bottom two long slits with dimensions of 10 μm by 200 μm are for speed detection. The other pinholes on the spatial filter are 10 μm by 20 μm (left) and 10 μm by 10 μm (right), for 3D image capturing with a pixel size of 2 μm and 1 μm in the $x$ direction, respectively.

Download Full Size | PDF

A 3D image Im can be generally described as

By using a cylindrical lens, the laser is diverged to form a light sheet with a height of 200–400 μm in the $y$ direction, depending on the design (10 or 20 pinholes in the spatial mask). In our treatment, we assume the light sheet is extended to infinity in the $y$ direction; thus the illumination intensity $I$ is only $z$- and $t$-dependent. We assume the scanning light-sheet illumination $I(z,t)$ can be described as a Gaussian beam,

With oversampling PMT signal readouts, the spatial resolution in the $z$ direction is diffraction-limited. The Gaussian beam waist is measured to be 0.73 μm.

To describe the relationship between the time-domain signal and the cell spatial information at position $x$, $y$, $z$, assumptions are made to simplify Eq. (4). The first assumption we made is to approximate $I(z,t)$ to be a delta function,

With the approximations above, when $y_{q+1}-y_q$ is larger than cell size (diameter), and cell projection is overlapped with $j$th pinhole,

Using this method, the system acquires one $yz$ plane image at each $x$ index, and all positions in the $yz$ plane have an identical PSF. Aided by the structured illumination, the light sheet further enhances the PSF of the detection system along the $z$ axis. As shown in Fig. 3(a), the PSF is calculated based on the diffraction model (ImageJ/Diffraction PSF 3D Plugin) with a 10X/0.28 objective lens. It has a long focal depth ($\sim 16\mathrm{\mu m}$) along the $z$ direction and $\sim 1.38\mathrm{\mu m}$ FWHM along the $y$ direction at its best focus. The light-sheet illumination profile, shown in Fig. 3(b), is described by a Gaussian beam with a measured FWHM of 0.73 μm. Consequently, the effective system PSF is sharp in both $y$ and $z$ directions, as shown in Fig. 3(c), in spite of slight blurring in the $y$ axis due to defocus [Fig. 3(d)]. The PSF at different $z$ positions is also measured by flowing 1 μm fluorescent beads [Fig. 3(e)]. The $z$ profile of the measured PSF is close to the light-sheet thickness, while its $y$ profile appears wider than the simulation result due to the blurring caused by the finite pinhole size.

 

Fig. 3. Simulated and measured PSF. (a) Simulated detection PSF with a 10X/0.28 objective lens in $yz$ plane; (b) simulated light-sheet cross section in $yz$ plane; (c) effective PSF in $yz$ plane, FWHM, $\mathrm{\Delta }y=1.38\mathrm{\mu m}$, $\mathrm{\Delta }z=0.73\mathrm{\mu m}$; (d) PSF variation along $y$ and $z$ axes; (e) experimental $yz$-plane image of a 1 μm fluorescent bead. The red curves in the profile plot are PMT raw signals; the black curves are fitted Gaussian curves. Scale bars, 5 μm.

Download Full Size | PDF

The laser used in the 3D-IFC system is a 1000 mW 488 nm laser (Coherent, Genesis MX SLM-488). The AOD (ISOMET, OAD948) is driven by a 200 kHz sawtooth voltage signal. The illumination objective lens is a 50X/0.55 Plan Apo Infinity Corrected Long WD Objective (Mitutoyo); the detection objective lens is a 10X/0.28, Plan Apo Infinity Corrected Long WD Objective (Mitutoyo) The light intensity signal detected by PMT (Hamamatsu, H10721-20) is first amplified by an amplifier (Hamamatsu, C9663) with a bandwidth from DC to 150 MHz, and then digitized by a digitizer (ADVANTECH, PCIE1840) with a maximum sampling rate of 125 MS/s per channel.

To precisely measure the speed of each cell for image reconstruction, two pairs of slits upstream and downstream of the pinholes are added to the optical spatial filter (Fig. 2). With the knowledge of cell travel speed across the detection zone and the scanning frequency of the light sheet, we can uniquely and precisely map each voxel in a 3D-IFC image to a specific time in the PMT output signal. 3D cell images can then be reconstructed from the time-domain signal. Figures 1(d)–1(f) shows a typical output of the 3D-IFC system. The detected time-domain multiparametric signals (multicolor fluorescence, FL1 and FL2, and SSC, light intensities) are synchronized with the applied voltage to the AOD, which scans the light sheet along the $z$ axis. At a 200 kHz scanning rate, a one-dimensional (1D) intensity profile in the $z$ axis is recovered from the time-domain signal within a time period of 5 μs [Fig. 1(d)]. During the $\sim 100\mathrm{\mu s}$ of cell travel between pinholes, $\sim 20$ periods of laser scanning are performed, and a $yz$-plane 2D image array is recovered from the PMT readout [Fig. 1(e)]. As the cell travels through the entire interrogation area, a stack of 2D $yz$-plane images are recovered, and the final 3D image can be reconstructed, with the ordinal pinhole number indicating the voxel’s $x$ position. In the example shown in Fig. 1(f), a 10-pinhole spatial filter produces 10 2D $yz$-plane images and a signal length of 1 ms, corresponding to a throughput of 500 cells per second. The bandwidth of the PMT and digitizer, 150 and 125 MHz, respectively, in this implementation, supports throughput of 6000 voxels per image acquired. Each cell in a volume of 20 μm by 20 μm by 20 μm, is captured as an image of 10 (or 20) $\text{pixels}\times 20\text{pixels}\times 50\text{pixels}$ with three color channels. The 3D image construction algorithm is realized in MATLAB according to the equations above. Due to the slight variance in the flow speed $v_C$ of cells, the original size of the 3D image of each cell can be slightly different. The captured cell image is then resized to $100\text{pixels}\times 100\text{pixels}\times 100\text{pixels}$ for 3D image analysis and quantitative measurements. 3D image batch processing is performed by ImageJ.

3. RESULTS

A. Cell with Fluorescent Beads

To demonstrate the cellular imaging capability of the 3D-IFC system, we imaged suspended single cells at a flow speed of 0.2 m/s. Figure 4 shows two-color fluorescence and unlabeled 90 deg SSC 3D images of mammalian cells. In Fig. 4, human embryonic kidney (HEK)-293 cells were stained with an intracellular carboxyfluorescein dye (CFSE) and bound with a random number of 1 μm fluorescent carboxylate-modified polystyrene beads. Figure 4(b) shows that while 2D bead images are overlapped, the 3D-IFC system resolves the exact number of particles from the reconstructed 3D images to address the occlusion issue, which is crucial for detecting localized and co-localized features [21,22]. In a flow system, cells and their internal structures are randomly orientated. As a result, 2D IFC systems may capture images from an unfavorable viewing perspective. Multiple perspectives can be achieved via 3D-IFC cell tomography to clearly resolve relative spatial relationships among features and count any spots and speckles correctly. The correct cell images with high information contents enabled by the 3D-IFC system not only accord with human visualization but also unleash the great potential for machine learning. Using the side-scatter dark-field imaging mode, the 3D-IFC generates a 3D spatial distribution of scattered light. It is known that refractive index variations will scatter light when the object is illuminated by visible light, and the refractive index distributions of the scattering regions are related to the intensity distribution of the scattered light. The 3D SSC images in Fig. 4 show the distribution of SSC light due to the refractive index ($n$) variations among the fluid (PBS, $n\sim 1.335$), the cells ($n\sim 1.3-1.6$) and the polystyrene beads ($n\sim 1.6$) [23,24]. As shown in Fig. 4(c), the intensity-based low-pass filtered SSC image indicates cell volume, and the high-pass filtered SSC image correlates with the fluorescent image of beads.

 

Fig. 4. Cells and beads imaged by the 3D-IFC. CFSE-stained HEK-293 cells bound with 1 μm fluorescent beads. (a) Recovered 2D $yz$-plane images and the assembled 3D surface-rendered view of CFSE fluorescence, bead, and SSC (bottom row); (b) representative 3D images of cells bound with beads (see also Visualization 1) and histogram detection events. The explicit relative position relationship in 3D space indicates that the particle counting in the 3D-IFC is independent of cell orientation. In the example of cell bound with four beads, occlusion in a specific perspective is a likely source of error for particle counting with 2D images. (c) Intensity-based processing of 3D SSC images. Left column, intensity histograms of 3D SSC image of the cell shown in (a). $P(x,y,z)$ is the position of 1 μm size bead determined using the 3D fluorescent image; within each bead position’s $\pm 1\mathrm{\mu m}$ area, the local intensity peak in 3D SSC image can be found. Scale bars, 5 μm. Flow speed 0.2 m/s. CFSE, intracellular carboxyfluorescein dye, Ex/Em: 488/517; bead, carboxylate-modified fluorescent microspheres, Ex/Em: 488/645; SSC, 90 deg SSC.

Download Full Size | PDF

B. $\gamma \mathrm{H}2\mathrm{AX}$ Foci Count in DNA-Damaged Cells

When ionizing radiation or cytotoxic chemical agents cause DNA damage in the form of double-stranded breaks (DSBs), the phosphorylated protein gamma-H2AX ($\gamma \mathrm{H}2\mathrm{AX}$) quickly forms foci at DSBs in a 1:1 manner [25]. With anti-$\gamma \mathrm{H}2\mathrm{AX}$ immunolabeling, foci count reflects the extent of DNA damage and the ability for DNA repair [26]. 2D microscopy imaging techniques and manual counting are used clinically today—a labor-intensive, low-throughput, and unreliable practice [27]. We demonstrated efficient and reliable foci counting for radiation-damaged cells using the 3D-IFC system with orders of magnitude higher throughput. We imaged immunolabeled $\gamma \mathrm{H}2\mathrm{AX}$ foci in CMK3 cells (a glioblastoma multiform cell line) after 6 Gy of gamma irradiation. Figure 5 shows representative cell images. The data show that the number of $\gamma \mathrm{H}2\mathrm{AX}$ foci is unrelated to the fluorescence intensity; thus, intensity-based measurements with conventional flow cytometry metrics poorly reflect DNA damage. In contrast, the 3D-IFC system enabled accurate and rapid analysis of the number of $\gamma \mathrm{H}2\mathrm{AX}$-positive foci.

 

Fig. 5. Fluorescent $\gamma \mathrm{H}2\mathrm{AX}$ foci imaged by the 3D-IFC. (a) Representative 3D images of irradiation-damaged glioblastoma CMK3 cells stained with CFSE and $\gamma \mathrm{H}2\mathrm{AX}$ antibody-conjugated PerCP/Cy5.5 and their two-color fluorescence 2D $yz$ plane merged image slices at $x=10\mathrm{\mu m}$. The high quality of the 3D images shows that the 3D-IFC is suitable for DNA damaged foci-related study. (b) Scatterplot of 917 detection events in the $\gamma \mathrm{H}2\mathrm{AX}$ intensity and foci count together with images of the cells within the marked regions (i)–(iv) in the scatterplot. The data show that foci count is unrelated to the fluorescence intensity from labeled $\gamma \mathrm{H}2\mathrm{AX}$; thus, intensity-based measurements with conventional flow cytometry metrics are unable to evaluate the extent of DNA damage. Scale bars, 5 μm; flow speed, 0.2 m/s; CFSE, intracellular carboxyfluorescein dye, Ex/Em, 488/517; PerCP-Cy5.5, DNA damage antibody-conjugated dye, Ex/Em, 490/677.

Download Full Size | PDF

4. DISCUSSION

Through several technological innovations, we have invented and demonstrated a 3D-IFC that captures and analyzes multiparameter 3D cell images at unprecedented throughput of around 500 cells/s. The 3D-IFC prototype features a cell flow speed of 0.2 m/s, a field of view of 20 μm, and a theoretical spatial resolution of 1 μm in all three axes. In the current implementation, our current system has 1, 2, and 4 μm resolution in the $z$, $y$, and $x$ axis, respectively, with a 10-pinhole design based on the Nyquist sampling theorem. The design has the flexibility to directly increase the spatial resolution, the field of view, and the 3D image capture rate through the change of spatial filter and the use of a higher flow rate. In the 20-pinhole design with half of the throughput, the resolution becomes 1, 2, and 2 μm in the $z$, $y$, and $x$ axis, respectively. A slower cell flowing speed or a higher light-sheet scanning rate can further improve the resolution in the $y$ direction towards the optical limit. Moreover, its information-rich 3D dark-field (SSC) image detection, coupled with 2D transmission image, offers possibilities for label-free assays. The 3D-IFC system offers unique capabilities for studies of many important biomedical characteristics, such as asymmetric division of T-cells into effector and memory cells [28], the secretory pathway of B cells, phenotype drug discovery [29,30], protein or receptor translocations [31,32], tracking of organelle formation or trafficking [33], chromosome structural aberrations [34], and 3D orientation and polarity, to name a few. The work also opens avenues for future integration with 3D image-based cell sorting [35,36] and machine learning for cell analysis.