Imaging fluorescence in moving cells is fundamentally challenging because the exposure time is constrained by motion-blur, which limits the available signal. We report a method to image fluorescently labeled leukemia cells in fluid flow that has an effective exposure time of up to 50 times the motion-blur limit. Flowing cells are illuminated with a pseudo-random excitation pulse sequence, resulting in a motion-blur that can be computationally removed to produce near diffraction-limited images. This method enables observation of cellular organelles and their behavior in a fluid environment that resembles the vasculature.
© 2013 OSA
Fluorescence microscopy enables the study of cells with molecular specificity and sub-cellular resolution [1,2]. Despite recent improvements in the brightness of fluorescent labels and in the sensitivity of image sensors, imaging is usually limited to stationary cells. Fluorescence imaging of cells traveling in flow could enable image based cell sorting and high-throughput microscopic characterization of extremely large numbers of cells [3–5]. In addition, non-adherent cells could be imaged in conditions similar to the blood stream where hydrodynamic forces are known to play an important role in cell morphology and function [6,7]. The temporal resolution of fluorescence imaging, however, is limited by a finite fluorescence emission rate . For a cell in fluid flow, this results in a trade-off between exposure time and motion-blur  and sets a limit on the achievable signal to noise ratio of the captured image.
Several methods in fluorescence microscopy have recently shown that it is possible to image inside stationary cells with a resolution that exceeds the diffraction limit . In addition to a spatial resolution limit, there is also a limit on temporal resolution that is governed by the intrinsic photophysics of fluorescence molecules. Photons are emitted by fluorophores at a rate fundamentally limited by their excited state lifetime, but in practice at a rate that is proportional to the laser excitation intensity. Frequently, the excitation intensity is kept as low as possible in order to reduce photoxicity and photobleaching . Long exposures are often needed to image weak signals or to collect high fidelity images. The time scale of the exposure, however, sets a limit on the amount of motion that can be tolerated by the object of interest. To collect a motion-blur free image, the object must move less than the minimum resolvable distance in the duration of the exposure time, therefore setting a motion-blur limited velocity (υl). Equivalently, for a given velocity, the exposure time must be less than the time it take the object to move a minimum resolvable distance, defined as (τl).
Fluorescence images of moving cells can be captured without motion-blur by using short camera exposure times, but the image quality suffers dramatically. Figure 1 shows brightfield and fluorescence images of a chronic myeloid leukemia (CML) cell lying stationary on a microscope slide. The CML cell has been labeled with the nucleic acid stain Syto16 that selectively stains the cell nucleus. The fluorescence images shown in Fig. 1(b)-1(c) are taken under the same excitation intensity, but Fig. 1(c) was captured with an exposure time that is 16 times longer than Fig. 1(b). The longer exposure clearly makes visible the otherwise almost invisible cell nucleus. As a simple metric to compare image fidelity, in this paper we define the image signal to noise ratio (SNR) as the ratio between the maximum signal value and the standard deviation of the background. If both the signal and background are photon shot noise limited, then the SNR is proportional to the square root of the exposure time . In agreement with this model, we find that the image in Fig. 1(c), with a 16 times longer exposure time, has an SNR that is improved by a factor of four.
Several methods have recently been developed that address limitations on high-throughput imaging of traveling cells by physically minimizing motion blur. One method is based on a time-delay-and-integration camera, which has been successfully implemented into an imaging flow cytometer . While this system’s performance is impressive, it requires precise control of the cell velocity in order to match the transfer rate of the electron readout. This restriction proves additionally limiting in cases where there is velocity dispersion either due to the fluid or the interaction between the cell and the channel. Another method that enables high throughput imaging in flow is to parallelize the microscope field of view [10–12]. In this way, the exposure time can be increased by the parallelization factor, but does not solve the problem of imaging fast moving cells.
2. Motion-blur encoding
Instead of physically minimizing motion blur, we investigate a computational method to eliminate it. Based on the assumption that the imaging process is linear and shift-invariant, motion-blur can be removed by deconvolution. However, a conventional exposure that is significantly longer than τl reduces high spatial frequencies in the collected image and the image deconvolution becomes ill-posed. Recent work in computational photography, however, has shown that by “fluttering” a camera’s shutter during the image acquisition , the blurred image retains its high frequency content and can be de-blurred with minimal loss in resolution. While to the best of our knowledge, motion-blur encoding has not been implemented in fluorescence imaging, a mask with a spatially modulated array of slits has been successfully used for the detection of fluorescently labeled particles in flow .
To implement a fluorescence imaging platform based on an engineered motion-blur, we have developed the system shown in Fig. 2. The excitation laser has a wavelength of 488 nm and an optical power of 50 mW that is focused to a 160 μm full-width-at-half-maximum spot size inside the microfluidic device. The camera is an Andor Neo scientific CMOS camera operated in global shutter mode. Physical encoding of the excitation light is performed by a custom made chopper wheel, shown in Fig. 2(a), which modulates the fluorescence excitation with a temporal code that is a pseudo-random sequence of 18 pulses. Two coded sequences are produced for every rotation of the chopper wheel. When the chopper is spun at 100 Hz, each pulse is 30 μs long, producing a total effective exposure time of 540 μs and the code has a 960 μs total duration. The individual pulse width sets the τl for the system, and assuming a spatial resolution of 1 μm, the motion-blur limited velocity of a single pulse is 33 mm/s. Cells flow through a microfluidic channel that is 30 μm wide and 8 μm deep. While velocity may vary from cell to cell, due to laminar flow in the microfluidic channel, the velocity of an individual cell stays constant during illumination and the trajectory follows the fluid streamlines.
Using the known code sequence, the velocity and the decoded image are found computationally using a parametric Wiener filter . The temporal code is first scaled by a constant that is based on an estimate of the particle velocity to transform the pulse temporal sequence into a spatial code, shown in Fig. 3(a). The spatial code is then convolved with a Gaussian function that corresponds to the point spread function (PSF) from a single pulse to produce the PSF of the entire system, shown in Fig. 3(b). We use a Gaussian that has an e−1 half-width of 400 nm in the flow direction and an impulse function in the height direction. The system PSF is then deconvolved from the measured motion-blur image using a Wiener filter implemented in MATLAB with the image processing toolbox. The Wiener filter takes as its input the system PSF, the motion-blur image, and a noise to signal power parameter (NSR). Both the synthesized point spread function and the motion-blur image were normalized to have a total integrated power of 1. In order to evaluate the optimal NSR and velocity scaling factor in the algorithm, we define an image quality metric based on the presence of deconvolution artifacts in the reconstruction. Deconvolution artifacts occur along the flow direction, indicated by the blue line in Fig. 3(c), when the NSR is too large or the scaling factor doesn’t match the particle velocity. The image quality metric is a ratio of the peak signal amplitude divided by the peak artifact amplitude. An iterative process is repeated on different velocity scaling factors to find a maximum in the image quality metric.
3. Imaging in flow
To explore image reconstruction using a coded excitation, we first imaged fluorescent beads traveling through a microfluidic channel. Figure 4 shows both raw and decoded images of 2 μm beads at four progressively decreasing excitation intensities. In all four cases, the bead is traveling at approximately 30 mm/s from left to right. The code modulates what would otherwise be a flat blur streak, approximately 30 μm in length. As can be seen from Fig. 4(e), even at an SNR below 1, the image can be reconstructed with reasonable fidelity. The code sequence, shown in Fig. 4(a), begins with an individual pulse. Because this pulse is isolated from the rest of the code by the two zeros that follow, we can use the beginning of the motion-blurred image to represent a single pulse exposure. This region of the motion-blurred image is denoted by a dashed line. We find that for all four excitation intensities, the reconstruction has an SNR that is approximately 10 times greater than that of the single pulse. This enhancement can be explained by a combination of the longer physical exposure and noise suppression from the deconvolution filter. While there is a trade-off between spatial resolution and background noise suppression in the implementation of the deconvolution filter, the longer effective exposure time enabled by the coded exposure increases signal strength without this trade-off.
By imaging fluorescently labeled cellular organelles, it is possible to observe their relative location and morphology in fast moving fluids. Figure 5 shows the nuclei of CML cells as they travel at increasing velocities through a microfluidic channel. The excitation sequence is the same as for the bead experiment. To resolve a 1.0 μm feature using an exposure time of 540 μs, the traditional motion-blur limited velocity (υl) would be limited to 1.9 mm/s. In contrast, Fig. 5(a) shows a CML cell traveling at 27 mm/s. Clearly, motion-blur is effectively removed, and intensity variations inside the nucleus can be clearly resolved. Interestingly, the velocity can be increased even past υl for a single 30 μs pulse, which corresponds to 33 mm/s. In this case, the deconvolution process must remove motion-blur from a single pulse in addition to the pulse sequence. As mentioned earlier, deconvolution of a single pulse is an ill-posed problem when the blur is much larger than the spatial resolution . However, we find that a moderate amount of motion-blur can be effectively removed. Figure 5(b)-5(d) shows cells traveling at 52, 80, and 103 mm/s, respectively, which continue to appear sharp and without blur, corresponding to as high as 54 times υl of the effective exposure time.
In addition, we have imaged peroxisomes of CML cells after transduction of a green fluorescent protein expressing construct. Peroxisomes are micron sized cellular organelles whose number and size is modulated as a function of the cell environment . Figure 6(a)-6(b) shows a raw and decoded image of peroxisomes in a cell traveling at 24 mm/s. Images were collected with longer 60 μs pulses and an effective exposure time of 1.08 ms. Using two peroxisomes as a two point resolution test, we observe resolution of peroxisomes spaced by 0.81 μm and 1.1 μm in the direction perpendicular and parallel to flow, respectively, shown in Fig. 6(c). This represents a conservative bound to the actual resolution, as the mean peroxisome size is measured to be 1.0 μm with 0.4 μm standard deviation, found by performing morphological analysis on the collected images. The resolution of the decoded image is slightly reduced in the flow direction due to the system PSF and Wiener deconvolution, but we still find that the decoded image is nearly diffraction limited in both directions. Deconvolved images of peroxisomes in two additional cells are shown in Fig. 6(d)-6(e).
We have presented a method that enables fluorescence imaging of cells traveling at high linear velocities in fluids. The image acquisition process is broken up into two steps, consisting of physical encoding and computational decoding of an engineered motion-blur. Further extensions could include imaging of cells on known but nonlinear trajectories or implementation of multiple orthogonal codes for multi-color resolution. In addition to imaging flow cytometry, we believe this method will make possible new studies on the dynamics of blood and cancer cells in conditions that resemble the vasculature. With this method, sub-cellular and organelle structure can be studied under hydrodynamic forces for the first time. The decoded images are nearly diffraction limited even though the effective exposure time can be more than 50 times greater than a traditional motion-blur limited exposure. By using a coded excitation sequence, we believe this system makes it possible to image fast moving cells with a spatial resolution that is comparable to what is obtained when imaging fluorescence in stationary cells.
This work was supported by the Rowland Junior Fellows program. We would also like to thank Adam Greengard for valuable discussions on coded imaging systems and Chris Stokes for help with electronics.
References and links
3. L. A. Sklar, Flow Cytometry for Biotechnology (Oxford, 2005).
4. T. R. Jones, A. E. Carpenter, M. R. Lamprecht, J. Moffat, S. J. Silver, J. K. Grenier, A. B. Castoreno, U. S. Eggert, D. E. Root, P. Golland, and D. M. Sabatini, “Scoring diverse cellular morphologies in image-based screens with iterative feedback and machine learning,” Proc. Natl. Acad. Sci. U.S.A. 106(6), 1826–1831 (2009). [CrossRef]
5. D. A. Basiji and W. E. Ortyn, Imaging and analyzing parameters of small moving objects such as cells. Amnis Corporation, assignee. US Patent 6211955, 2000–03–29 (2001).
11. E. Schonbrun, A. R. Abate, P. E. Steinvurzel, D. A. Weitz, and K. B. Crozier, “High-throughput fluorescence detection using an integrated zone-plate array,” Lab Chip 10(7), 852–856 (2010). [CrossRef]
13. R. Raskar, A. Agrawal, and J. Tumblin, “Coded exposure photography: motion deblurring using fluttered shutter,” ACM SIGGRAPG 25(3), 795–804 (2006). [CrossRef]
14. P. Kiesel, M. Bassler, M. Beck, and N. Johnson, “Spatially modulated fluorescence emission from moving particles,” Appl. Phys. Lett. 94(4), 041107 (2009). [CrossRef]
15. R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image Processing Using Matlab (Gatemark, 2009).
16. R. A. Saleem, B. Knoblach, F. D. Mast, J. J. Smith, J. Boyle, C. M. Dobson, R. Long-O’Donnell, R. A. Rachubinski, and J. D. Aitchison, “Genome-wide analysis of signaling networks regulating fatty acid-induced gene expression and organelle biogenesis,” J. Cell Biol. 181(2), 281–292 (2008). [CrossRef]