Fluorescence array detector for large‐field quantitative fluorescence ...

Cytometry 16:206-213 (1994)

0 1994 Wiley-Liss, Inc.

Fluorescence Array Detector for Large-Field Quantitative Fluorescence Cytometry' K.D. Wittrup? R.J. Westerman, and R. Desai Department of Chemical Engineering, University of Illinois, Urbana, Illinois Received for publication July 9, 1993; accepted January 26, 1994

A prototypic fluorescence array detector (FAD) has been designed and constructed which is capable of quantifying single-cell fluorescence emissions from a statistically significant population of cellsized objects (over lo3) on a solid substrate. The system is comprised of a cryogenically cooled CCD, 50 mW air-cooled argon ion laser, and optics that image a large (1 x 1 cm) field at 1:l (no magnification). The CCD is effectively treated as a two-dimensional array of 2.7 x lo5 independent 20 x 20 pm photodetectors,

Flow cytometry allows quantification of single-cell fluorescence while detecting sufficient numbers of cells to obtain a statistically significant estimate of the population distribution. The flow cytometer measures objects serially at high speed (11,14) and is widely used, although it has several limitations, including inability to use samples mounted on solid substrates; clogging of flow assemblies due to debris (8);inability to repeat a measurement on the same cell; and photodestruction of the sample (3). Image cytometry allows measurements of samples on solid substrates, based on either a scanning excitation beam, a movable stage, or a combination of the two (1,4,6-9,13). While these devices allow one to rescan the sample repeatedly, intense excitation tends to cause photodestruction of the sample, and a small Sample of cells is present in a single field of view of a typical fluorescence microscope objective. Galbraith et al. (5) have recently described an image cytometer for multicolor measurements on a large number of cells imaged a t relatively low magnification (20 x ), based on a conventional epifluorescence compound microscope with a computer-controlled scanning stage. Acquisition of data from a population of lo3 cells requires approximately 2 h (5). We have designed and constructed a simple imaging cytometer called a fluorescence array detector (FAD) that achieves high sample throughput by imaging a large field (1 x 1 cm) a t 1 : l (no magnification) onto a

with each cell-sized object imaged across only a few CCD pixels. Algorithms have been developed for focusing, image segmentation, shading correction, and noise rejection; performance data for the FAD with fluorescent calibration beads are presented. The FAD is a simple alternative to microscope-basedimaging cytometry, allowing large-field imaging without a scanning stage. 0 1994 Wiley-Liss, Inc. Key terms: Image cytometry, CCD, large field imaging

cryogenically cooled CCD detector, using two inverted large format fil.0 camera lenses. The CCD is utilized essentially as a two-dimensional (2D) array of independent photodetectors, with a few detectors per cell. Since the cell images are approximately the size of a CCD pixel (20 x 20 km), spatial sampling is well below the Nyquist frequency and essentially all single-cell morphological information is lost. Simultaneous quantitative measurements of the single-cell fluorescence intensity from a large population of cells are possible with the FAD without a movable stage or scanning beam, making the FAD system relatively simple, inexpensive to construct, and mechanically robust. An air-cooled 50 mW argon ion laser provides low irradiance excitation, minimizing photodestruction of the sample, allowing the FAD to achieve high sensitivity through long integration times, consistent with a theoretical examination by Mathies et al. (10) of fluorescence detection optimization, which suggests that lower excitation irradiance over a longer observation time yields a higher signal to noise ratio. PC-based software has been developed for noise correc-

'This work was supported by an NSF Presidential Young Investigator Award (NSF BCS-90-57677) to K.D.W. 'Address reprint requests to K.D. Wittrup, Department of Chemical Engineering, University of Illinois, Urbana, IL 61801.

I .

FLUORESCENCEARRAYDETECTOR

CCD

Signal to image processing

Detector

1x1 f/l .0 50 mm

”

Interference

50 mW

camera lenses

FIG.1. Schematic diagram of an FAD. A 1 x 1cm sample is imaged without magnification on a CCD chip, which is treated as an array of independent photodetectors. The sample is illuminated by a laser beam incident at Brewster’s angle (33”)to minimize reflections. Specifications of individual components are given in the text.

tion, image analysis, and data interpretation. The FAD represents a simple means of acquiring quantitative cytometry data on surface-affixed cells without a substantial investment in equipment beyond the CCD camera itself.

RESULTS A schematic diagram of the FAD is shown in Figure 1. The sole moving part is the sample stage, which is used for focusing. The system is comprised of a CCD, a n air-cooled argon ion laser, two high numerical aperture camera lenses, a focusing micrometer, and a Dell 325 (Austin, TX) computer for data acquisition and analysis. The sample and CCD are mounted on a vibration isolation table. The sample is directly illuminated by a n expanded laser beam and the fluorescent cell population is imaged without magnification directly onto the CCD surface. An epi-illumination configuration is unnecessary due to the large (2 cm) working distance between the collection optics and the sample. An epiillumination configuration with the camera lens optics was found to result in increased noise due to fluorescence and scatter from the internal dichroic mirror (data not shown).

CCD Fluorescence emissions are detected with a cryogenically cooled Photometrics charge coupled device (CCD chip PM512, CCD Model CH 210, Tucson, AZ). A CCD was chosen due to its highly linear response, geometric stability, high dynamic range, and low noise characteristics (16). The PM512 sensor 2D array consists of

207

516 x 516 pixels, each measuring 20 pm2. The CCD has a readout rate of 50 kHz and 14 bit gray scale resolution.

Imaging Optics The collection and imaging optics meet two design criteria: large-field imaging and maximization of light collection without compromising image quality. In order to image a large (1cm2) field a t low magnification, two high-speed camera lenses (Hughes-Leitz, Ottawa, Canada) were used in a head to head configuration, spaced 6 in. apart at the lens faces. The f / l lenses have a 50 mm focal length and up to a 40 mm format size. In the head to head configuration, light from the sample is collected and imaged to infinity by the first lens. The collimated light is then filtered by two 7 cavity interference filters (534 df 22 by Omega Optical, Brattleboro, VT, and 545 df 45 by Chroma Optical, Brattleboro, VT). Two emission filters are used to cover any pinhole defects that may be present in either filter. After passing through a n aperture to reject scattered light and unfocused rays, the second camera lens refocuses the sample image onto the CCD detector face. Focusing is accomplished with a movable stage attached to a micrometer z-translator (Oriel, Stratford, CT) with 0.25 pm resolution. Rough focusing is performed by eye, with a n image of approximately 100 fluorescent 6 pm (diameter) polystyrene beads (Polysciences, Warrington, PA). Fine focusing is achieved by adjusting the stage z-translation micrometer to minimize the number of pixels per bead image with intensity above a user-supplied threshold. An example of a typical minimization of this parameter by focusing is shown in Figure 2 and indicates a n approximate depth of focus of 20-40 pm. An image of a resolution test target (USAF Resolution Test Target 1951, Melles Griot, Irvine, CA) is shown in Figure 3, as a crude measure of image quality. The smallest features distinguishable by eye are the lines in group 4, element 4, which are approximately 22 pm in width, which is the size of the CCD pixels. Aliasing due to misalignment of the target features and the CCD array is inevitable, but it is apparent that the optical train resolves features approximately the size of the detector discretization. An aperture between the lenses determines the overall collection efficiency of the optical train, rejects outof-focus scattered light, and also affects image quality, flatness, intensity, and the size of the imaged field as shown in Figure 4. For the majority of the images described, a n internal aperture of 1in. is used. Although the lenses nominally are capable of fil imaging (N.A. = 0.5), measurements of the light collection efficiency described below indicate less efficient collection, which is a n inevitable consequence of very large-field imaging. However, the low noise characteristics of the CCD enable sensitive fluorescence detection despite such inefficient light collection.

208

WITTRUP ET AL.

10000 , 8000

5 4 U

6000

2000 0

0 100

0

-50

50

100

z FIG. 2. Focusing of FAD by minimization of number of Pixels Per bead image. A field of 6 pm (diameter) beads (PolYsciences)was imaged repeatedly, changing the stage height between each The average number of pixels per object (120 objects in field) was calculated for the image and plotted vs. stage position. A depth of field of 20-40 pm is inferred from the width of the minimum.

FIG.3. FAD optical train imaging resolution is greater than CCD pixel discretization. A negative USAF Resolution Test Target 1951 was backlit and imaged with the FAD optics; a blowup of the central portion of the image is shown. Features on the image which are approximately the size of a CCD pixel (20 x 20 pm) can be resolved (group 4, element 4 lines are 22 pm in width).

Illumination In order to obtain quantitative fluorescence intensity data, the excitation source must be spectrally pure, spatially uniform, and temporally stable. To meet these criteria, vertically plane polarized light from a 50

100

200

300

400

500

Diagonal Distance across field (pixels) FIG. 4. Effect of aperture on image intensity and field flatness. A flatly illuminated field [a flashed opal diffuser disk (Oriel) over the aperture of an integrating sphere (Oriel)] was imaged by the FAD, and intensity vs. pixel location along the diagonal is plotted for varying internal aperture sizes. Apertures larger than 1 in. did not alter the image,

mW wavelength tunable argon ion laser (Omnichrome 532 AP, Chino, CA) is first filtered with a 488 nm laser line filter (Omega Optical). The TEM 00 (Gaussian profile) beam is then expanded by a series of four cylindrical lenses into a n elliptical spot (Melles Griot). After passing through a final dichroic sheet polarizer (Melles Griot) to ensure p-polarization, the beam impinges on the sample plane at 33" (Brewster's angle) to minimize reflection of excitation light from the sample substrate. The approximate irradiance at the sample plane is 1 mW/cm2. The laser is run under light control mode to ensure that excitation intensity at the sample plane is temporally stable, which is essential when obtaining long exposures of weakly fluorescent objects. A shading correction is applied to correct for inhomogeneous sample illumination and imaging aberrations. The laser's Gaussian beam Drofile Droduces uneven illumination, and light from the middle of the field is collected by the optical train more efficiently than light from the edge of the field. A uranyl glass slide with a fluorescence spectrum similar to fluorescein is imaged for the shading correction (gift of D. Kaplan).

Image Processing Hardware and Software Image analysis was performed on a Dell 325 PC (25 MHz, 8 MB RAM). The Photometrics camera was interfaced using the National Instruments (Austin, TX) general purpose interface bus. All image analysis software was developed using Turbo C + + (Borland, Scotts Valley, CA) along with libraries from the Heap Expander (The Tool Makers, Santa Cruz, CA) to allocate expanded and extended memory for data storage above the 640 kb limit (a single digitized image from the CCD is over 500 kb). Software was written for fo-

FLUORESCENCE ARRAY DETECTOR

5500

0

1

2

3

4

209

I

I

1

I

I

I

5

FITC molecules (x lO-g)/pixel FIG.5. FAD light collection efficiency. To test collection efficiency, rectangular capillary tubes were filled with a series of FITC solutions ranging from to lo-" M and imaged for 0.1 s ( + ) 0.5 s (01,and 1.0 s (A),The rate of photoelectron accumulation is plotted against fluorophore concentration, and the least-squares line (r2 = 0.99) indicates that 5.8 x photoelectrons are collected per fluorophore per second.

cusing, object identification, noise correction, shading correction, and data display. The focusing and noise correction algorithms are described more fully below. Software has been developed for image segmentation utilizing a user input threshold criterion to discern objects from background, determined empirically by examination of the mean pixel intensity in a n area free of fluorescent objects. Assigning pixels identified in the segmentation process to a n object is based on 8 neighbor connectivity. After object identification, aggregates are discarded on the basis of size.

Light Collection Efficiency The light collection efficiency of the FAD was determined using flat capillaries (0.1 mm light path, Vitrodynamics, Rockaway, NJ) filled with a series of dilutions of fluorescein isothiocyanate (FITC) solutions (lop5to lop8M). The capillary volume was determined to be 4.6 -t 0.1 p1 from the difference in mass before and after filling with deionized water. The dye-filled capillaries were suspended in air at the focal plane, imaged for 0.1,0.5, and 1 s, and corrected for shading errors a s described above. Average intensity values were taken from pixels along the length of the capillary. The results are shown in Figure 5. From the least-squares fit to the data, 5.8 x photoelectrons are detected per fluorescein molecule per second. Given the expected losses from reflections and filter band pass width, this collection efficiency is consistent with a n effective aperture of f/4. To confirm these sensitivity measurements, 6 pm (diameter) flow cytometer calibration beads (Flow Cytometry Standards Corporation, Triangle Park, NC) with 5,000 mean equivalent soluble fluorescein

FIG.6. Photobleaching kinetics in the FAD. Polysciences fluorescent calibration beads were mounted under a coversh in PBS and SlowFade reagent (Molecular Probes, Eugene, OR) and illuminated continuously by a 50 mW argon ion laser beam at 488 nm. Four consecutive images of 0.6 s duration were taken a t time points during the illumination and the average total intensity for each bead was determined. Results for a typical bead are shown, and a least-squares curve fit (r2 = 0.99) indicates a photobleaching rate constant of 0.002 min-', corresponding to a bleaching half-time of 5.6 h.

(MESF) were imaged. The number of photoelectrons detected per bead per minute in a series of 15 min exposures is 18.6, for a detection rate of 6.2 x photoelectrons per fluorophore per second. This is in close agreement with the collection efficiency measured with FITC-filled capillaries.

Photobleaching Since the FAD illumination irradiance is very low (approximately 1 mW/cm2), photobleaching is very slow compared to fluorescence microscopy or flow cytometry. Photobleaching was measured with 6 pm (diameter) calibration beads (Polysciences) suspended in phosphate buffered saline (PBS) under a coverslip sealed with enamel nail polish. Four 0.6 s exposures were taken a t time points during 2 h of continuous illumination; the shading-corrected integrated f luorescence intensity of a typical bead a t the center of the field is plotted in Figure 6. The beads at the center of the field bleach fastest due to the Gaussian profile of the excitation source. A least-squares fit of the data to first-order bleaching kinetics gives a photodestruction rate of 0.002 min-l. Thus, the FAD photodestruction half-life is approximately 6 h, a photobleaching rate 8 orders of magnitude slower than in a flow cytometer sK1 for FAD compared to 3,500 sP1in flow cytometry (1511.

Noise Sources In order to detect weakly fluorescent objects with the FAD (=lo3 fluorophores), it is necessary to integrate for time periods up to a n hour. Some noise components of the signal (substrate fluorescence, CCD thermal electrons, stray light, and spurious events) accumulate

210

WITTRUP ET AL.

Table 1 FAD Noise Sources

Contribution Dark noise Stray room light Stray laser reflections Fused silica coverslip (0.015 in. thick) Schott black glass (UG-1) Titanium oxide coverslip Soda lime microscopic slide

Rate (electronsimin) 0.015 0.012 0.71 9.90 6.20 89.9 1,000

with increasing time of signal collection, while others (readout amplifier noise, relative photon shot “noise”) do not. A consideration of the different sources of FAD noise allows determination of the potential signal to noise ratio of this instrument. Contributions of each of the above noise sources to the FAD signal were measured and are summarized in Table l . Each noise source and procedures for its abatement are described below. Readout amplifier noise results from discretization of charge, in addition to noise from the preamplifier circuitry. Since preamplifier noise is incurred only when the array is read out, i t is not a function of the time of signal acquisition. Therefore, the ratio of the fluorescence signal to preamplifier noise can be arbitrarily increased by increasing exposure time. Preamplifier noise can be estimated from the difference between successive bias images (acquisitions of zero time duration). Pixel-by-pixel image-to-image variation appears symmetrically distributed around zero, with a roughly Gaussian distribution. This noise source can be minimized by slowing the rate a t which the array is read out. At the PM512 maximum readout speed of 50 kHz, the preamplifier noise contributes a standard deviation of 7.7 photoelectrons upon successive bias readouts of a given pixel, and when readout is slowed by use of a larger gain (Photometrics CGAIN 2101, preamplifier noise is reduced to 4.0 photoelectrons. Dark noise results from thermionic emission of electrons within the silicon lattice of the CCD chip due to thermal motion of the lattice. These “dark” electrons are indistinguishable from photoelectrons, and follow Poisson statistics. Dark noise is minimized by cryogenic operation (-13OoC), by turning off the chip preamplifier during long periods of signal acquisition, and by operating the CCD in multipinned phase (MPP) mode. Under these conditions, the rate of generation of dark electrons is 0.015 electrons/pixel/min. Stray light reaching the detector can result from a number of sources: excitation photons leaking through the emission filters; excitation photons causing components in the optical train to fluoresce; and room light in the emission pass band. Both emission filter leakage and optical train fluorescence are minimized by adjusting the angle of the p-polarized excitation light to Brewster’s angle, thereby minimizing reflection of the excitation beam into the collection optics. To minimize

room light in the emission pass band, the instrument is located in a photographic dark room and shrouded with black velvet. The rate of acquisition of stray light photoelectrons was determined by opening the shutter with the laser turned on, without a sample in the focal plane. After subtracting the known rate of generation of dark electrons, the stray light noise component of the signal is 0.72 electronsipixellmin. All detectors suffer from photon “shot noise” due to the Poisson nature of photon emission from a fluorescent object. The variance of a Poisson random number is equal to the mean, so the coefficient of variation is inversely proportional to the square root of the mean. Shot noise is minimized by collecting a larger number of photons (e.g., longer integration times). A long period of signal acquisition would be required in order to completely overcome shot noise limitations a t the present FAD excitation intensity, but use of a 10 W water-cooled argon ion laser could boost illumination excitation intensity by 3 orders of magnitude. It has been estimated that approximately lo5 photons are emitted from a fluorescein molecule before destruction by photobleaching, on average (15). Since 0.1% of the emitted fluorescence photons are detected by the FAD, approximately lo2 photons will be detected from a fluorophore before photodestruction. Thus, the lowest attainable shot noise for a single fluorophore gives a coefficient of variance of 10%. The substrate used for cell immobilization contributes noise to the signal, both from substrate fluorescence and stray reflections. Standard microscopic slides (soda lime glass) and coverslips (titanium oxide) generate a large amount of green fluorescence when illuminated at 488 nm (1,000 and 89.9 electrons/ pixelis, respectively). Ultraviolet (UV) grade fused silica coverslips (dimensions 38 x 22 x 0.38 mm) were obtained from Commercial Crystal (Naples, FL), and samples inserted between two such coverslips. The rate of generation of fluorescence from these sample substrates is 9.9 electronslpixelimin. Although the observed signal is referred to here a s fluorescence, it should be noted that the photons may be generated a t the substrate surface by a n inelastic scatter process or a contaminant introduced in the cleaning process. In any case, the signal from the substrate is not reduced by insertion of additional emission interference filters, indicating that the detected photons are not merely reflected excitation light at 488 nm. Clearly, the dominant noise contribution is substrate fluorescence. Various cleaning procedures were ineffective a t further reduction in this noise source. Given the measured efficiency of light collection and the summed contribution of all known noise sources, the noise signal is equivalent to 1,800-2,900 fluorescein molecules per pixel. Substantial improvement in this noise contribution would be possible with improved substrate material. The CCD dark noise is equivalent to only 4 fluorophore molecules per pixel and the stray light contribution is approximately 200 fluorophores

211

FLUORESCENCEARRAYDETECTOR

10

per pixel. It is to be expected that a focused effort a t eliminating the noise contributions from stray light and substrate material could result in a n FAD capable of detecting on the order of 100 fluorophore molecules.

Discrimination of Spurious Events With Dixon Rank Order Statistic A potentially troublesome source of FAD noise results from spurious events that appear as bright spots in images (even when the shutter is closed), which may be caused by cosmic particles, radioactive decays, or uncharacterized chip-related events. These spots occur with a frequency of approximately 1 per minute per field, with a n average size of 14.3 pixels per event. These events are difficult to discriminate from cells in a particular image by eye, given the lack of single-cell morphological information. To discern spurious events from real objects, we exploit their rare frequency of occurrence. Six consecutive images of the sample are taken, and successive cell or bead images vary in intensity according to Poisson statistics. However, spurious events generally appear a t a given location in only one of the six images. The spurious events can be identified a s outliers using a Dixon rank order statistic (2). The Dixon statistic is defined as the ratio of the difference between the highest value and its nearest neighbor and the difference between the highest and lowest values, when the six intensity values for a given pixel are sorted in rank order. For a normally distributed random variable, it can be shown that samples with a Dixon statistic greater than 0.698 represent a n outlier with 99% confidence. Although this cutoff is derived for Gaussian statistics, a similar number could be derived for Poisson statistics. Empirically, a cutoff value of 0.698 provides effective discrimination. If the Dixon criterion for a given pixel indicates the occurrence of a spurious event, the outlier is discarded and its value replaced with the mean intensity value of that pixel in the other five images. This method effectively eliminates spurious noise from FAD images, as shown by comparing the pixel intensity histograms for a half hour exposure (with the shutter closed) before and after spurious event rejection (see Fig. 7). The Dixon criterion does not rely on estimation of population moments, which are skewed by outliers in a small sample. For example, a large outlier in six observations alters the estimates of the mean and standard deviation (S.D.) sufficiently that the outlier will fall within 1 S.D. of the estimated mean. It should be noted that Dixon statistics discriminate spurious events well only when they are rare. If two events occur at the same location in two different images, the scheme described does not discard them. Sample Statistics The FAD images a large field of cells (1 x 1 cm). Figure 8a shows a n image of 6 pm (diameter) calibration beads (Polysciences) and the enlarged view in Fig-

8

m

6

a,

.-x

Q

a

4 2

0 0

2000

4000

6000

8000

10000

10

b 8

m

6

a,

.-x

Q

*

4 Fixel intensity

2

0 0

2000

4000

6000

8000

10000

Pixel intensity FIG.7. Discrimination of spurious events with Dixon rank order statistics. Six 30 min dark exposures were taken, during which time spurious events accumulated on the array. a: A histogram of pixel intensities for one of the images is shown. b: The Dixon rank order statistic was used to reject those pixels which possessed intensities which were not relatively constant throughout the six images; the resulting histogram of pixel intensities after correction is shown.

ure 8b shows the relative size of bead images and pixels. Beads were suspended in propanol for better dispersion, spread on a microscopic slide, and allowed to dry. The beads were imaged for a series of 10 exposures. The images were corrected for shading errors using uranyl glass a s a flat-field image and corrected for spurious events by Dixon statistics. The first image in the series was thresholded using a user input threshold level and image segmentation was performed. Integrated bead intensities were corrected for stray light by subtracting the local background value from each pixel in the object. The object file was then corrected for doublets and higher level aggregates based on object size (analogous to gating on forward angle light scatter in flow cytometry). Each bead occupies a n average of 6 pixels in the image, most likely due to a combination of

212

WITTKUP ET AL.

30

X

C

X

X

x

X

x x

lo

5

object intensity

140

I

I

120 100

80

60 40 20 0

object intensity

Magnification of boxed area imperfect focusing and pixel discretization aliasing effects. The histogram and dot plot for the image in Figure 8a are shown in Figure 8c. The coefficient of variance (CV) of measured bead fluorescence intensity is 12.9%, although flow cytometric measurements of beads from the same stock yield a CV of 4%. Single bead integrated intensity in five consecutive images exhibits a CV of less than 2%. Single bead integrated intensity in five subsequent im-

FIG.8. FAD image and histogram of 6 +m (diameter) polystyrene beads (Polysciences). a: A negative representation of a n FAD image of fluorescent calibration beads. Six 10 s exposures were taken and corrected for shading, and Dixon statistics were applied to eliminate outliers. b: A 3.5-fold enlargement of the boxed area in a, to highlight the relative size of bead images and CCD pixels. c: A composite histogram from 5 independent fields, with a total of 1,791 beads. The CV of the singlet peak is 12.9%.

ages moving the sample between images yields a CV of 7%. Thus, imperfect flat-field shading corrections are responsible for a CV larger than that obtained by flow cytometry. Inexpensive cylindrical lenses are used in the beam expander of this prototypic instrument and subtle interference patterns are visually apparent in the illumination field. It is likely that small shifts in these patterns are responsible for the spatial inhomogeneity measured.

FLUORESCENCEARRAYDETECTOR

DISCUSSION The FAD is a n image cytometer capable of acquisition of flow cytometry-type data from a large number of cells immobilized on a solid surface. While flow cytometry measures events in series, the FAD is a massively parallel instrument. The FAD is a simple, flexible, robust instrument with only one moving partthe focusing stage. The quality of each component of the FAD may be chosen to meet the criteria of a particular application-if extreme sensitivity is not required, then the low noise characteristics of a cryogenically cooled CCD may be rejected in favor of a more economical thermoelectrically cooled CCD. If samples with relatively high levels of fluorescence emission are to be observed, then a lower intensity and less expensive arc lamp or tungsten filament light source may be used. The quality of the emission interference filters can also be chosen to be consistent with the application and by changing emission filters multicolor fluorescence measurements could be performed. Our motivation in constructing the FAD is to perform quantitative single-cell fluorescence in situ hybridization measurements on recombinant yeast populations. However, many other potential applications exist. Detection of rare fluorescent cells in a large population is particularly straightforward with the FAD, since a large number of cells may be observed and the sparse distribution of rare fluorescent cells ensures simple image segmentation. An example of such a n application would be detection of human immunodeficiency virus (H1V)-infected cells in blood cell populations (12). Polymerase chain reaction (PCR)-amplified in situ hybridization can detect HIV-infected cells, “but a large number of microscopic fields need to be surveyed by a trained observer to demonstrate a rare affected cell” (12). The FAD is particularly well suited to applications such as this. The rapidity of analysis should allow high throughput for a n FAD device in a clinical setting. The work presented here demonstrates the FAD’S potential sensitivity. Currently, the limiting noise source is substrate fluorescence. This could be reduced in the future through the use of different excitation sourceldye combinations. The current argon laser is wavelength tunable and therefore capable of exciting dyes a t longer wavelengths (i.e., rhodamine or CY3 a t 514 nm). Using a n infrared (IR) laser with longer wavelength cyanine dyes would increase signal due to the CCD’s higher sensitivity in the IR (16) while lowering both substrate and cellular autofluorescence. One of the most significant difficulties to overcome with FAD imaging cytometry is the preparation of cell

213

samples with adequate separation to allow image segmentation in the absence of any single-cell morphological information. For microbiological or cell suspension culture samples, this is a straightforward matter of dispersion and adhesion. Distribution of cells on a n evenly spaced grid would allow observation of individual cells without requiring excessively dilute dispersions to reduce doublets. Such cell patterning is possible with self-assembled monolayer lithography on gold substrates (Kumar, personal communication). However, for tissue samples or surface-dependent cell culture, discrimination of individual cells may be difficult. In summary, the FAD is a simple device for imaging cytometry which should be useful for a number of research and clinical applications as a low-cost alternative to microscope-based cytometers.

LITERATURE CITED 1. Allison DC, Mayall BH, Levin J : Comparison of absorption measurements of DNA stain content by utilizing video and scanning image cytometers. Cytometry 9573-578, 1988. 2. Barnett V, Lewis T: Outliers in Statistical Data, 2nd Ed. John Wiley & Sons, Inc., New York, 1984, pp 171-172. 3. Bock G, Hilchenbach M, Schauenstein K, Wick G Photometric analysis of antifading reagents for immunofluorescence with laser and conventional illumination sources. Histochem Cytochem 33(7):699-705, 1985. 4. Burger D, Gershman R Acousto-optic laser-scanning cytometer. Cytometry 9:lOl-110, 1988. 5. Galbraith W, Wagner MCE, Chao J , Abaza M, Ernst LA, Nederlof MA, Hartsock R J , Taylor DL, Waggoner AS: Imaging cytometry by multiparameter fluorescence. Cytometry 12579-596, 1991. 6. Jaggi B, Poon SS, MacAulay C, Palcic B: Imaging system for morphometric assessment of absorption or fluorescence in stained cells. Cytometry 9566-572, 1988. 7. Jovin TM, Arndt-Jovin DJ: Luminescence digital imaging microscopy. Annu Rev Biophys Biophys Chem 18:271-308, 1989. 8. Kamentsky LA, Kamentsky LD: Microscope-based multiparameter laser scanning cytometer yielding data comparable to flow cytometry data. Cytometry 12:381-387, 1991. 9. Lee BR, Haseman DB, Reynolds C: A digital image microscopy system for rare-event detection using fluorescent probes. Cytometry 10:256-262, 1989. 10. Mathies RA, Peck K, Stryer LB: Optimization of high-sensitivity fluorescence detection. Anal Chem 62:1786-1791, 1990. 11. Melamed MR, Lindmo T, Mendelsohn ML (eds): Flow Cytometry and Cell Sorting, 2nd Ed. Wiley-Liss, Inc., New York, 1990. 12. Patterson BK, Till M, Otto P, Goolsby C, Furtado MR, McBride LJ,Wolinsky SM: Detection of HIV-1 DNA and messenger RNA in individual cells by PCR-driven in situ hybridization and flow cytometry. Science 260:976-979, 1993. 13. Shack R, Baker R, Buchroeder R, Hillman D, Shoemaker R, Bartels PH: Ultrafast laser scanner microscope. J Histochem Cytochem 27:153-159, 1979. 14. Shapiro HM: Practical Flow Cytometry, 2nd Ed. Alan R. Liss, Inc., New York, 1988. 15. Taylor DL, et al. (eds):Applications of Fluorescence in Biomedical Science. Alan R. Liss, Inc., New York, 1989, pp 129-140. 16. Wang YL, Taylor DL (eds): Methods in Cell Biology, Vol29, Part 16. Academic Press, Inc., New York, 1989.