We demonstrate high throughput gigapixel fluorescence microscopy with a microlens array. We show, for the first time to the best of our knowledge, the use of a parallelized microscopy system to image samples in micro well plates. We image centimeter-scale regions of 384-well micro well plates at 1.72 μm resolution at a raw pixel throughput of 25.4 Mpx/s. Taking into account the fact that about half the well plate area consists of the plastic support region between wells, this corresponds to a sample pixel throughput of 13.2 Mpx/s, more than double that of the commercial state-of-the-art at the time of writing. Fluorescent imaging of tissue samples through coverslips is also demonstrated.
©2013 Optical Society of America
Optical microscopy is a workshorse technique in modern biological research . Many areas of inquiry, such as drug discovery, are greatly facilitated by extremely large data sets . These are becoming ever more commonplace, enabled by continual improvements in the cost:performance ratio of computing power and digital storage capacity. These huge data sets often come in the form of large-scale fluorescence micrographs. Commercial units, termed high throughput automated microscopes, are tasked with imaging centimeter-scale cell cultures at micron resolution. Typically, the cell cultures are grown in an arrayed holder called a microwell plate. A microwell plate is made up of an array of isolated cell culture wells, each typically housing a distinct cell culture. The area of a single well in a 384-wellplate is 11.56 mm2, giving a total imaging area of 44.4 cm2 . At a spatial sampling frequency of 0.5 μm/pixel, this corresponds to 17.7 gigapixels (Gpx) per wellplate. Using a five megapixel (Mpx) camera, which represents a typical value used in microscopes at the time of writing, one could construct such an image from approximately 3500 fields-of-view (FOVs). Indeed, today’s automated microscopes build up a large image by sequentially imaging and stitching together thousands of FOVs. The serial nature of this approach makes it inherently slow, due to the fact that there is temporal overhead associated with the mechanical scanning process, namely the time needed to move the stage, allow it to settle, and refocus the image. From the point of view of imaging throughput, these steps represent dead time, during which the image sensor remains idle.
There are currently many efforts to push imaging into the Gpx regime at both larger length scales [4, 5], and in microscopy [6–8]. In microscopy, these large FOV techniques use scanned focal spot arrays to parallelize imaging. These techniques have highly scalable architectures, meaning that they can take advantage of the improvements in computer and camera technologies that are currently taking place in order to increase imaging throughput.
Here, we demonstrate imaging of samples loaded into a micro well plate, the first time this has been reported for any parallelized microscopy system, to the best of our knowledge. This is significant because, at the time of writing, micro well plates are used extensively in high throughput drug screening labs . These labs have made considerable investments in robotic systems which handle well plate samples. Therefore, imaging techniques that are compatible with well plates are highly desirable for high throughput, large FOV microscopy. We also image a tissue microarray and a fluorescently labeled section of mouse kidney, demonstrating the sensitivity and resolution of our system with biologically relevant samples. This work represents a considerable scaling-up of our previous implementation of fluorescence imaging with microlens arrays . We increase the imaging area from 0.3 cm2 (0.55 cm × 0.55 cm) to 12.96 cm2 (3.6cm × 3.6cm), with a total increase in raw pixel throughput from 4 Mpx/s to 25.4 Mpx/s. We furthermore demonstrate water-immersion imaging through a coverslip, unlike our previous demonstration of microlens array imaging through air and without a coverslip . These coverslips comprise either the bottom glass of a wellplate, or a standard #1.5 coverslip for the tissue imaging experiments.
2. Experimental setup and fabrication
We now describe the optical setup (Fig. 1) in detail. A green laser beam (λ = 532nm) is sent through a microscope objective (50 × and 10 × for wellplate and tissue imaging experiments, respectively). The resultant divergent beam is then reflected by a long pass dichroic mirror (cutoff λ = 550nm) and converted into a beam that diverges more weakly by a tube lens (f = 200mm). The diverging beam is clipped both at the dichroic mirror and at the tube lens aperture, so that only the central portion of the beam illuminates the sample; a small amount of beam divergence is necessary because the microlens array is larger than the tube lens aperture. The excitation uniformity over a 3.6 × 3.6 cm region of interest (ROI) at the sample when using the 50 × objective is 15% (center to edge). A refractive microlens array (pitch 100μm, f = 345μm in water) is located at a distance of 200 mm from the tube lens. The microlens array focuses the laser beam into an array of focal spots at the sample plane. The beam divergence must be small enough for the off-axis aberrations experienced by focal spots at the periphery of the ROI to be also small. In our case the beam divergence is <7.5°, small enough so that off-axis aberrations are insignificant. The microlens array is mounted on a tip/tilt stage (Edmund Optics), enabling it to be adjusted to be parallel to the sample. The sample is placed at the focal spot array plane on a 2-axis closed loop piezoelectric stage (Newport NPXY200SG) that scans the sample over a 110 × 110 μm square at the focal spot array plane. Water is used as an immersion liquid between the lens array and sample in order to increase focal length. It also has the benefit of improving the resolution compared to what would be obtained with a microlens array of the same focal length operating in air. The fluorescence excited at each focal spot is collected by its respective microlens and relayed through a long pass filter (cut on λ=575 nm) to a CMOS camera (Basler acA2000-340km, 2048 × 1088 pixels) via a combination of the tube lens and a single lens reflex camera lens (SLR, Nikon 28-80mm, f/3.5). The camera sensor is located one focal length behind the SLR lens, and records an image of the fluorescence as seen at the back of the microlens array. In all experiments presented here, the SLR lens focal length is set to be 30 mm, meaning that the image on the camera sensor has a magnification of 0.15× (30mm/200mm). One can consider the microlens array to function as an array of point scanning microscope objectives, with the camera performing the task of a detector array.
The raw image at the camera is an array of bright spots (Fig. 1(i)), each corresponding to the fluorescence excited and collected by one microlens. By recording the intensity relayed by a given microlens as the sample is scanned, an image of the portion of the sample directly under that lens may be obtained. The camera is operated at 200 frames per second (fps), and the raster scan line rate is set to an integral number of camera frames to avoid image shearing. Data is streamed from the camera to the computer where it is losslessly compressed and continually saved to disk for offline processing. The rate at which the software can compress the data is the limiting factor for the pixel throughput reported in this paper.
The microlens array is fabricated via a reflow molding process [10, 11]. A 695 × 695 rectangular array of cylindrical posts (11 μm tall, 94 μm diameter, 100 μm pitch) is written onto a 100 mm diameter silicon wafer in SPR220-7.0 photoresist using a direct write lithography system (Heidelberg μpg-501). After development, the wafer is placed on a hotplate at 140°C for 1 minute to melt the photoresist pillars into smooth lenses. The microlens height after melting is 18.3 μm. A negative of the photoresist master is then replicated in polydimethylsiloxane (PDMS) . Finally, the photoresist template is replicated in a polymer (Norland Optical Adhesive NOA 61, n = 1.565) on a 50 × 75 mm microscope slide using the inverse PDMS mold as a stamp. The microlens array spans an area of 50 × 69.5 mm and contains ~347,000 elements. A micrograph of a portion of the microlens array is shown in Fig. 2(a). The microlens array has a focal length of 141.5 μm in air and 345 μm in water, the latter yielding a theoretical numerical aperture (NA) of 0.18.
3. Experimental results
3.1 Imaging resolution
In order to determine the resolution of our system, we image sub-resolution fluorescent beads, i.e. having diameters (500 nm) that are smaller than the spot size. We use a raster scan step size of 385 nm and a pixel integration time of 4.5 ms. The image of a sub resolution bead provides a good approximation to the point spread function (PSF) of a fluorescence imaging system . We fit a two dimensional Gaussian to the images of each of ten beads dried onto a microwell plate. The mean full-width-at-half-maximum (FWHM) of the bead images is 1.72 μm +/− 0.08 μm, and indicates the resolution of our system. This is slightly worse than would be expected of a diffraction limited lens 0.18NA lens (1.49 μm). This is due to the spherical aberration inherent when producing microlenses via resist reflow .
3.2 Well plate imaging
A 384-well microwell plate (Figs. 3(a) and 3(b)) is loaded with fluorescent beads (Nile red, Invitrogen, 5 μm diameter) in water and then left to dry. This sample is then imaged as described above with an integration time of 1 ms per pixel, and raster scan step size of 540 nm. A microscope objective with a magnification of 50 × (Fig. 1) is used in order to create a relatively flat excitation field. The images formed by the microlenses are stitched together using a custom written MATLAB script to produce a 6600 × 6600 pixel image of each well. The stitching procedure uses the 10 μm overlap between each pair of FOVs to compute the relative signal gain and pixel offset between FOVs, thereby correcting for possible fabrication imperfections. The overlapping areas of each pair of FOVs (e.g. FOV 1 and FOV 2) are then multiplied by the vignetting functions V1(u) = (u/b)3/2 (applied to FOV 1) and V2(u) = 1-V1(u) (applied to FOV 2), where u is distance from the FOV 1 border, and b is the width of the 10 μm overlap region in pixels. Note that for any two overlapping regions, the sum of the vignetting functions is unity over the entire overlapping region. After vignetting, each FOV is translated by the previously calculated pixel offset and the vignetted signal in the overlap regions is summed, yielding nearly seamless stitching of thousands of FOVs .
The image of the well plate given in Fig. 3(a) is shown in low resolution, since the image at full resolution could not of course be readily shown in one figure panel. The photograph of Fig. 3(b) shows the region on the well plate from which the image is obtained. We now consider the raw pixel throughput, i.e. not accounting for stitching. There are 360 × 360 microlenses, and each microlens produces a 110 × 110 μm image with 540 nm pixels, i.e. 203 × 203 pixels. The raw data therefore corresponds to 2032 × 3602 = 5.34 Gpx. This data is captured in a 3.5 minute scan, meaning that the raw pixel throughput is 25.4 Mpx/s.
Images of typical wells are shown in Figs. 3(c)-3(f), and are produced by the stitching process described above. We simultaneously image 64 wells. Each well is imaged by an array of 34 × 34 microlenses that correspond to a region on the camera sensor with an extent of 96 × 96 pixels (~8 pixels / microlens). Counting only the wells, and not the plastic support regions between them, the scan records 2.78 Gpx of sample data (6600 × 6600 px image/well × 64 wells) and a sample pixel throughput of 13.2 Mpx/s after stitching overhead is taken into account. This throughput is more than double that of commercial state-of-the-art systems such as the ImageXpress Micro from Molecular Devices, which achieves 4-5 Mpx/s .
When imaging the 64 well sample, the image on the camera takes up 976 × 976 pixels, corresponding to 360 × 360 microlenses. Thus, we use less than half of the available camera sensor area and microlenses. This is because the desktop computer employed is not capable of streaming the data sufficiently fast for the entire sensor and entire microlens array to be used. For a 64 well scan with a raster scan step size of 540 nm, the compressed 8-bit data from the camera is ~44 GB. This limitation could be overcome with a faster image compression algorithm, more RAM or a solid state hard drive array to store the uncompressed data.
3.3 Tissue imaging
We use our lens array to image a hematoxylin and eosin (H&E) stained tissue microarray (TMA) slide (USA Biomax, BLSC1501), mounted on a microscope slide and sealed with a #1.5 coverslip. TMA slides are used in biological research for high throughput gene expression studies, often for cancerous tissues [17,18]. Typically H&E staining is used for analysis with brightfield transmission microscopy. Here, however, we image the eoisin dye using our fluorescence system, because it provides a convenient demonstration of fluorescent imaging in biological samples. We image a TMA slide in one 3.5 minute scan at a raster scan step size of 540 nm and an integration time of 4.5 ms per pixel. Typical images of 1.1 mm diameter tissue slices are shown as Figs. 4(b) and 4(e). Zoomed in crops of the tissue cores in Figs. 4(c), 4(d) and 4(f) show the level of detail captured in each image. We note that because a TMA slide is small compared to the 8 × 8 well region of the microwell plate imaged as described above, we do not take full advantage of the pixel throughput of our technique. A larger TMA with a denser array would be more appropriate with our setup and could see sample pixel throughputs exceeding even that of the well plate sample.
We also image a coverslip-sealed mouse kidney section stained with AlexaFluor 568 Phalloidin (Invitrogen FluoCells Slide #3). The sample has an extent of 1 × 0.6 cm (Fig. 5(a)). Actin filaments in the glomeruli are visible in Fig. 5(b).
Because both tissue samples have weaker fluorescent emission than the microsphere sample, we use an objective with a magnification of 10 × to expand the beam, rather than one with a magnification of 50 × . This means a larger portion of the original beam is reflected by the dichroic mirror into the excitation path, resulting in more power at each excitation focal spot. It also results in a more uneven (Gaussian) illumination profile at the microlens array, which is apparent in Fig. 4(a). Imaging larger tissue samples would require a higher power laser spread out over a larger area, perhaps coupled with more efficient beam shaping optics such as a flat and collimated beam generator. At the current illumination intensity we have not observed any photobleaching effects in our system.
We have demonstrated high throughput gigapixel fluorescence imaging using a water immersion refractive microlens array. A resolution of 1.72 μm is achieved, and imaging is performed through standard #1.5 thickness coverslips and microwell bottoms. We achieve a sample pixel throughput of 13.2 Mpx/s, which is more than double that of the commercial state-of-the-art at the time of writing. We have shown that that even though the camera has a modest sensitivity, i.e. it is not an EMCCD nor a sCMOS camera, biological samples with typical fluorophore densities can readily be imaged with sufficiently short dwell times to enable high throughput imaging. In future implementations, larger tissue samples will be imaged, requiring the use of higher laser powers.
Coupled with sensitive EMCCD or sCMOS cameras, our technique would be capable of imaging at low fluorophore densities while maintaining high throughput. Additionally, the detection efficiency and resolution could be improved by fabricating higher NA refractive or diffractive microlenses [19–21].
Due to its high throughput, we expect that our system would greatly benefit the biological imaging communities. We furthermore note that its architecture is such that the continual advances in image sensors and computing power that are taking place at the time of writing are anticipated to lead to similarly dramatic advances in imaging throughput.
This work was funded by the National Science Foundation (NSF, grant number ECCS1201687). Fabrication work was carried out at the Harvard Center for Nanoscale Systems, which is supported by the NSF.
References and links
3. Ibidi μ-Plate 384-well specifications sheet, http://ibidi.com/fileadmin/products/labware/plates/P_8840X_Plate_384well/IN_884XX_384.pdf, accessed 24 Nov 2012.
4. D. J. Brady, M. E. Gehm, R. A. Stack, D. L. Marks, D. S. Kittle, D. R. Golish, E. M. Vera, and S. D. Feller, “Multiscale gigapixel photography,” Nature 486(7403), 386–389 (2012). [CrossRef] [PubMed]
5. O. Cossairt, D. Miau, and S. K. Nayar, “Gigapixel computational imaging,” IEEE International Conference on Computational Photography, 1–8 (2011).
6. G. Zheng, X. Ou, and C. Yang, “Towards giga-pixel microscopy,” CLEO 2012. Optical Society of America, Washington, DC, 1–2.
7. S. O. Isikman, A. Greenbaum, W. Luo, A. F. Coskun, and A. Ozcan, “Giga-Pixel Lens Free Holographic Microscopy and Tomography Using Color Image Sensors,” PLoS ONE 7(9), e45044 (2012). [CrossRef] [PubMed]
9. S. Michael, D. Auld, C. Klumpp, A. Jadhav, W. Zheng, N. Thorne, C. P. Austin, J. Inglese, and A. Simeonov, “A Robotic Platform for Quantitative High-Throughput Screening,” Assay Drug Dev. Technol. 6(5), 637–657 (2008). [CrossRef] [PubMed]
10. P. H. Nussbaum, R. Volkel, H. P. Herzig, M. Eisner, and S. Haselbeck, “Design, fabrication and testing of microlens arrays for sensors and microsystems,” Pure Appl. Opt. 6(6), 617–636 (1997). [CrossRef]
11. A. Schilling, R. Merz, C. Ossmann, and H. P. Herzig, “Surface profiles of reflow microlenses under the influence of surface tension and gravity,” Opt. Eng. 39(8), 2171–2176 (2000). [CrossRef]
12. M. V. Kunnavakkam, F. M. Houlihan, M. Schlax, J. A. Liddle, P. Kolonder, O. Nalamasu, and J. A. Rogers, “Low-cost, low-loss microlens array fabricated by soft-lithography replication process,” Appl. Phys. Lett. 82, 1152–1154 (2003).
13. R. W. Cole, T. Jinadasa, and C. M. Brown, “Measuring and interpreting point spread functions to determine confocal microscope resolution and ensure quality control,” Nat. Protoc. 6(12), 1929–1941 (2011). [CrossRef] [PubMed]
14. F. T. O’Neill and J. T. Sheridan, “Photoresist reflow method of microlens production part II: analytic models,” Optik (Stuttg.) 113(9), 405–420 (2002). [CrossRef]
16. Molecular Devices ImageXpress Micro product website, http://www.highthroughputimaging.com/screening/imagexpress_micro.html#apps, accessed 24 November 2012.
17. J. Kononen, L. Bubendorf, A. Kallioniemi, M. Bärlund, P. Schraml, S. Leighton, J. Torhorst, M. J. Mihatsch, G. Sauter, and O. P. Kallioniemi, “Tissue microarrays for high-throughput molecular profiling of tumor specimens,” Nat. Med. 4(7), 844–847 (1998). [CrossRef] [PubMed]
18. P. Capodieci, M. Donovan, H. Buchinsky, Y. Jeffers, C. Cordon-Cardo, W. Gerald, J. Edelson, S. M. Shenoy, and R. H. Singer, “Gene expression profiling in single cells within tissue,” Nat. Methods 2(9), 663–665 (2005). [CrossRef] [PubMed]
20. 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] [PubMed]
21. E. Schonbrun, P. E. Steinvurzel, and K. B. Crozier, “A microfluidic fluorescence measurement system using an astigmatic diffractive microlens array,” Opt. Express 19(2), 1385–1394 (2011). [CrossRef] [PubMed]