A single platform image cytometer for resource-poor settings to monitor disease progression in HIV infection

Sample Preparation for Image- and FCM

Blood from random-selected HIV-negative patients was collected in K₃EDTA vacutainer blood collection tubes and processed within 10 h after draw. Fresh blood specimens were supplied from the MST Hospital, Enschede, The Netherlands. Note that patients not diagnosed with HIV are defined as HIV-negative patients.

For image cytometry, 100 μl EDTA anti-coagulated whole blood was mixed with 10 μl of CD3 EasySep (ES) Ab cocktail (StemCell Technologies, Vancouver, BC) and 10 μl CD4 R-Phycoerythrin (PE). After 15-min incubation at room temperature, 5 μl ES magnetic nanoparticle solution was added and mixed. After 10-min incubation at room temperature, the sample was diluted with system buffer (Immunicon, Huntingdon Valley, PA) to a final volume of 410 μl and mixed. About 320 μl of the sample solution was transferred into an analysis chamber (Immunicon) with inner dimensions of 30 × 2.7 × 4 mm³. Separate titration experiments were performed to determine the optimal concentration of both immunomagnetic and fluorescent labels, the incubation time, and the magnetic separation time (12). For CD4 and CD8 tests, we used 3 μl of CD3 Ferrofluid (Immunicon) for magnetic separation, and 5 μl CD4 PE and 2 μl CD8 PerCP (BD Biosciences, San Jose, CA) for fluorescent detection.

For FCM (TruCount method, BD), blood samples were processed using a lyse-no-wash method according to recommendation of the manufacturer (BD) (13) and analyzed on a FACSCalibur flow cytometer (BD).

CD4 and CD8 T Lymphocyte Enumeration Method

After incubation with CD3-magnetic particles and labeling with CD4-PE and CD8-PerCP, the blood sample is injected into a sample chamber that is placed in the field of a magnetic yoke (14, 15). Here, the CD3⁺ T lymphocytes, being CD4⁺ and CD8⁺ T lymphocytes, are extracted from the sample and pulled up toward the analysis surface (glass) of the sample chamber while other cells move to the bottom of the sample chamber under the influence of gravity, as shown schematically in Figure 1. Fluorescence emitted from a well-known area of the analysis surface is sequentially imaged for each fluorophore, and the number of CD4⁺ and CD8⁺ T lymphocytes in the area is determined by image analysis. Magnetically separated cells derive from a precisely known volume of the chamber. Consequently, the absolute number of CD4⁺ and CD8⁺ T cells per unit volume, and therefore CD4/CD8 ratio are determined. Although CD4⁺ T lymphocytes stain brighter with CD4 than do monocytes, the presence of dim CD4⁺ T lymphocytes makes identification of monocytes by adjusting the fluorescence detection levels of the image cytometer more difficult. We used PE and PerCP as fluorescent labels, because these two fluorophores can be excited by the same light source, (see Imaging Optics section), and there is a minimal overlap between their emission spectra, minimizing the cross-talk that may occur between them.

Illumination Configuration

The cytometer should give a high signal-to-noise ratio (SNR) for images of separate cells, so that they can reliably be counted. This means that the intensity of illumination should be both high and homogeneous over the analysis surface. A nonhomogeneous illumination would cause a nonuniform fluorescent signal and a nonuniform background, leading to problems in the image analysis. The illumination scheme implemented in our cytometer consists of four 5 W, 470 nm, Luxeon V Star LEDs (LXHL-LE5C, Lumileds, San Jose, CA) mounted symmetrically on top of the magnetic yoke. Each LED unit requires a typical operating voltage of 6.84 V at 700 mA and outputs a luminous flux of 48 lm (Lambertian radiation pattern).

We used a fiber coupling lens with a focal distance of 16 mm (Roithner Lasertechniek, Vienna, Austria) to collimate the output of each LED onto the upper surface of the sample chamber. A lens holder for Luxeon LEDs (Fraen Corporation, Reading, MA) assembles LED and lens, and this assures proper relative positioning of these components. A 470AF50 excitation filter (Omega Optical, Brattleboro, VT) was positioned in front of the lens to reduce the LED output overlapping with the fluorescence of PE, hence, contributing on reduction of the background. This filter blocks to optical density >4.0 for the wavelength range 350–440 nm and 500–800 nm, as measured with a spectrophotometer (Shimadzu UV-2401PC, Shimadzu, Kyoto, Japan). All these optical components were aligned and packed together using 25-mm shrink tubing (Farnell In One, Leeds, UK). A schematic representation of an assembled LED module is shown in Figure 3. The four legs of each module are heat-staked onto a metallic plate, which serves also as heat sink. In this way, an excellent optical assembly with high mechanical stability is obtained. The four LED modules are positioned such that the optical axis makes an angle of 45° with the plane of the stage. Given the illumination geometry and the magnetic yoke configuration, the 45° was optimal to obtain a maximum light intensity with relatively good uniformity at the top of the sample chamber. The distance between each LED module and the top of the chamber can be properly adjusted. The optimum distance was found to be ∼20 mm. For this distance, each LED module delivers ∼40 mW/cm² at the top of the chamber, as measured with an optical power meter (model 840, Newport Spectra-Physics BV, Utrecht, The Netherlands). The overall light intensity distribution on the measuring area, where beams of four LED units overlap, has a coefficient of variation (CV) of <10% as determined with the WinCamD beam diagnostics (Gentec-EO, Québec, QC).

Imaging Optics

The number of pixels used to image one cell is an important parameter in the design of the instrument. To obtain good discrimination of hot pixels, the number of pixels per cell should be maximized. This will lead to a decrease in the SNR when the light intensity per pixel is compared versus the noise in surrounding pixels. In addition, the accuracy is increased because the resolution increases, and it will become easier to separate two neighboring cells. On the other hand, the larger the amount of pixels used per cell, the fewer cells can be imaged on the CCD surface, which will lead to poor sampling statistics. Therefore, the number of pixels per cell was optimized to obtain the largest SNR with a minimum number of pixels per cell.

Fluorescently labeled cells, lying at the upper surface inside the chamber, are imaged onto the CCD camera using a 10× microscope objective (NA 0.2) (Lomo Optics, Germantown, MD) mounted on a stage that enables adjustment of the focus. A sample chamber with fluorescent beads is used to determine the focus upon which the adjustment is locked. In the cytometer, we use the objective to realize a magnification of 6.3×. The CCD covers a field of view (FOV) of 1.09 × 0.73 mm² of the upper surface. Fluorescently labeled CD4⁺ T lymphocytes are randomly located on the object plane and imaged as spots of ∼25 pixels, which was found as a reasonable compromise between SNR and statistical variation. The lymphocytes within one FOV come from a volume of 0.80 μl of the labeled blood sample (FOV multiplied by the depth of the chamber). For each sample, cell counts were performed for three FOV areas and the counts were averaged. For blood with a CD4⁺ T lymphocyte count of 50/μl, this procedure should yield a statistical (Poisson) variation of <10% of the count.

Factors contributing to the background signal are autofluorescence from the labeled and nonlabeled cells at the upper surface of the chamber and from the serum with the red blood cells in the sample chamber, as well as reflected and stray LED light and room light. Use of a proper emission filter will suppress part of these factors and minimize average background signal. In our image cytometer, the PE fluorescence, emitted by labeled CD4 cells, was filtered by a 595AF60 emission filter (Omega Optical). This filter is blocked to optical density >4.0 for the wavelength range of 400–550 nm and 670–800 nm, being sufficient for achievement of a high SNR (see Instrument Characterization section). For PerCP, a 695AF55 emission filter (Omega Optical), having similar performance as the 595AF60, was used. A filter changer was mounted to switch between these two filters, enabling capturing of PE and PerCP signals.

The emission spectrum of a 470 nm LED, as measured with a monochromator (HR460; Jobin-Yvon, Paris, France), containing a blazed holographic grating with 1,200 grooves/ml, is presented in Figure 4, together with the characteristics of the excitation and emission filters, and the absorption and emission spectra of PE. The spectrum of the LED matches well with the absorption spectrum of PE. A higher yield could be achieved by using a LED with an output wavelength of 505 or 530 nm, also available from Lumileds. Then, however, a larger interference of the LED light with the emission of PE occurs, because the emission spectrum of these LEDs is broad (spectral width at half maximum intensity 30 nm). This results in a higher background in the images. As will be shown in Instrument Characterization section, the choice of the excitation wavelength is conditioned by the SNR of cell images.

CCD Camera

The noise present in lymphocyte images arises from photon shot noise of the emitted (auto)fluorescent light and stray light that passes through emission filter, and from noise generated by the CCD camera (dark current noise, readout noise, and thermal noise). An additional contribution to the noise stems from the variance in sensitivity of separate pixels of the CCD chip. Therefore, it is important to use a CCD camera with low noise characteristics and minimal sensitivity variation. The CCD camera used (ST-402ME; SBIG, Santa Barbara, CA) was designed for astronomical purposes, where faint details of dim astronomical objects have to be determined. In analogy, we have to image the weak immunofluorescence of labeled cells. This camera is offering a high performance at relatively low costs. The imaging CCD chip (Kodak KAF-0402ME imaging CCD; Kodak, NY) has 765 × 510 pixels (9 μm × 9 μm), and combines a large well capacity (∼100,000 e⁻) with a low dark current (1 e⁻/pixel/s at 0°C) and a low readout noise (17 e⁻ RMS). The camera has a 16 bits A/D converter. Owing to the implementation of a microlens array over the pixels, it has a high quantum efficiency (∼75% at the peak of PE emission). Dark image subtraction can be performed for images acquired with exposure times longer than 1 s. The camera can operate from a 12 V battery or other unregulated 12VDC source.

Single Board Computer and Touchscreen

For data collection and analysis, we use a single board computer (SBC). Compared to a normal computer or a laptop, the SBC is an inexpensive solution that can easily perform overall cytometer control. Furthermore, its size, weight, and power requirements allow development of stand-alone platforms. The cell images are downloaded from the CCD camera to the SBC. This is a Micro PC with a VIA Eden™ ESP6000 667 MHz processor (EES-3610, Evalue Technology, Taiwan); a 256 MB SDRAM SODIMM memory (SimpleTech, USA) and a 2.5″ hard disk (Toshiba, Japan). The Micro PC is a low-power fanless embedded system (operating with 5 V-DC at 5 A) with dimensions of 220 mm × 65 mm × 147 mm and weighs 1.7 kg. An 8.4″ LCD-TFT touchscreen monitor (B084SN03 V2, AU Optronics, Taiwan) with a JTHC084DRA1 touchscreen controller (AU Optronics) is connected to the SBC. The touchscreen is powered by a 12 V external voltage.

Housing and Power Supply

The total power required for cytometer operation is 55 W, from which about 50% is used by SBC and touchscreen. To minimize power consumption, the LEDs are controlled via a USBREL8LC relay card (Quancom, Germany) and operate only during the image acquisition.

The cytometer needs 5 V for the SBC, 6.8 V for the LED units, and 12 V for the CCD camera and LCD-TFT touchscreen. A home-built power unit contains one Yuasa NP 7-12 12 V rechargeable battery with a capacity of 7 Ah (Yuasa, Japan), that is charged by a LM12010-3T 12V 1A battery charger (MEC, Hong Kong). Two LT1083 voltage regulators (Linear Technology, USA) with an output 5 V and 6.8 V are used. The power supply is connected via a five-pole connector to the cytometer. In spite of the small battery, the current prototype runs for about 2.5 h, without the need for recharging the battery. In this time, 100 cell counts can be determined. An external car battery can be connected to the power unit to operate in case of power loss. A 30-Ah battery would then last for about 10 h operation.

A home-built preliminary aluminum housing of 25 cm × 25 cm × 20 cm accommodates all components of the cytometer, except for the power unit. The optical parts, including CCD and magnetic yoke, are positioned at the front part of the instrument. A schematic layout of this section is shown in Figure 5A. The user inserts a sample chamber from the front into the magnetic yoke. The SBC is mounted on the rear vertical plate, which serves as heat sink. The power unit, built in a separate housing, is mounted on the outside of the rear plate. The touch screen is positioned on top of the cytometer at an angle of 45°. A photograph of the cytometer is shown in Figure 5B. Total weight is 15 kg; instrument without power unit weighs 10 kg.

Image Analysis

To obtain the number of CD4⁺ and CD8⁺ T lymphocytes per unit volume, we analyze the fluorescence images using a dedicated algorithm written in ImageJ, a public- domain Java image processing software package (NIH, MD) (16). The algorithm consists of the following steps.

Background subtraction: Corrects for nonuniform illumination, variation in the sensitivity of the CCD pixels, and presence of dust in the optical system.

Thresholding: Transforms the grey-scale image to a binary image. Here, all the objects (cells) that have higher intensity than a preset threshold are assigned as 1, and the rest of the objects that have lower intensity (background) are assigned as 0.

Watershedding: Separates cells that are close together (17). The watershedding transformation works by replacing all foreground (cell) pixels with a grey value that is proportional to their distance from the nearest background pixel. This is the so-called Euclidian distance map. From this, the centers of the cells are calculated by finding the ultimate eroded points, i.e. the points that are equidistant from the cell edges. These points are then dilated until they meet another dilated object, and there a segmentation line is drawn.

Counting: Enumerates all objects assigned as 1 (cell).

Separation: An additional step that first estimates the average area of a cell in the image and then divides the total area of overlapping cells by this average. This step is necessary for images with a high cell count to allow a better separation between overlapping cells.

All steps of the algorithm are implemented in a macro that enables an automated image analysis. Analysis of one image takes less than 10 s in the SBC. To determine the quality of the overall procedure of acquisition of an image and its analysis, we follow Young et al. (18) and define a quality index (QI) of an image with cells as QI = (〈I_cell〉 − 〈I_backg〉)/〈σ_noise〉, where 〈I_cell〉 and 〈I_backg〉 are the average fluorescence intensity of the lymphocytes and the background from several representative areas in the image, and 〈σ_noise〉 is the average standard deviation of the total noise present in the image. Note that the maximum lymphocyte intensity for each area is taken as 〈I_cell〉. Using this definition, the QI can serve to determine how well an image was analyzed. We use the term QI instead of SNR as it also contains the accuracy due to the image analysis algorithm. A selection of the cells counted in the original image, serving as a qualitative assessment of the counting procedure applied, is provided from the algorithm.

In Figure 6, we illustrate the image analysis. It shows a typical image of CD4 T lymphocytes, magnetically selected using CD3-ES, and fluorescently labeled with CD4-PE. The results of background correction, thresholding, and watershedding are also presented. In this case, apparently, all cells have been identified by the image analysis procedure. Note that this image corresponds to a blood sample having around 1,500 T-lymphocytes/μl, illustrating that the algorithm works well, even for high CD4 counts. However, in general, we estimate that counts can reliably be made in the region of <1,500 cells/μl, and above this concentration deviations may occur (see also Evaluation of the CD4 and CD8 T-Cell Enumeration section).

Software Shell

A software shell (LabView, National Instruments, Austin, TX) that integrates all hardware and software into a functional and simple user interface was developed. The measurement procedure is started by inputting the patient info and the sample number. Next, the exposure time needed for acquiring a sample image is set. After a dark image and sample images for PE and PerCP have been captured, the dark image is subtracted from the sample images. Finally, the image analysis takes place, and a report that contains the patient info, the CD4 count and CD4/CD8 ratio, and the images of the blood sample is presented. The whole procedure from image capture to the cell count result takes less than 2 min.

Instrument Characterization

During the development of the image cytometer, we built and tested several configurations of LED-filter-collimator configurations. The QI of representative cell images, being the guiding criterion for prototype development, was evaluated for each configuration. A summary of five different configurations that were explored is presented in Table 1.

The QI improved from 23 in Configuration 1 with two 1 W 470 nm LEDs, to 69 in Configuration 5 with four 5 W 470 nm LEDs. This is a result of several factors, such as a higher excitation power, better collimating optics, introduction of an exciter filter that reduces LED output overlapping with the PE fluorescence, and use of a more efficient emission filter. In Configuration 5, because of the high QI, the amount of CD4-PE used for immunofluorescent labeling was decreased from 10 to 5 μl. Configuration 5 also gives the best result with respect to background uniformity, with a CV of 5% (along the field profile as shown in Fig. 7) vs. a CV of 41% in Configuration 1. This improvement is due to use of better collimating optics and proper alignment of LED modules. Using this configuration, it was possible to detect the BD QuantiBrite beads, bearing ∼5,956, 26,653 and 69,045 PE molecules, with a QI > 30 (exposure time of 10 s). In this case, extrapolation of the experimental data indicates that it might be possible to discriminate beads bearing a few hundred of PE molecules.

Next, the influence of the QI on the accuracy of cell counting was investigated by analyzing images with known number N of CD4⁺ T lymphocytes. Blood samples with a CD4 count that varies from about 50 to 500 cells/μl were used. We selected this range because it seemed most relevant for AIDS monitoring in resource-poor settings. From each image, several images with QIs varying from about 5 to 40 were generated by adding random noise to the original image. Next, these images were analyzed with the algorithm defined in Image analysis section, and a relative counting error was determined for each image. This error is defined as (|N′ − N|)/N × 100%, where N′ is the number of counted cells. Figure 8 shows that the relative error of cell counting decreases with increasing QI. For QI > 30, the error that results from the simulations is less than 2%. Considering that, in actual measurements, the QI of the cell images we obtain is usually better than 30, we expect that the image analysis inaccuracy is less than 2%. These simulations also illustrate the dynamic range of the image analysis algorithm in the clinically relevant range of 50–500 cells/μl. Calculations were also performed with images of fluorescent beads that have a similar size as T lymphocytes (Fig. 8). These simulations confirm the results achieved with images of CD4⁺ T lymphocytes.

We compared these results with simulations performed with generated cell images using Matlab software (Mathworks, Natick, MA). These generated images resemble the experimental settings, such as magnification, used for cell imaging and size of the CCD chip. The same image analysis procedure as for the real cell images was applied, and the error was estimated for various number of cells/μl. A good agreement between the results obtained with the generated and real cell images was found, confirming the accuracy achieved with the cytometer. Moreover, analysis of both generated and real cell images indicates that the sum of counting error and statistical (Poisson) variation is less than 10% in the range of 50–1,700 cells/μl.

Evaluation of the CD4 and CD8 T-Cell Enumeration

The performance of our SP ICM was evaluated by measuring samples of HIV-negative patients and comparing the results with the TruCount method. To test CD4 count, samples from 45 patients were analyzed. Parallel blood samples of each patient were processed in both instruments. The correlation coefficient and the regression line for results of our image cytometer and the flow cytometer are shown in Figure 9A (R = 0.98, P < 0.0001; slope 0.97). The correlation coefficient of 0.98 indicates a good agreement. The slope of 0.97 indicates that the deviation from FCM is 3%. The Bland–Altman plot (19) that compares absolute CD4⁺ T lymphocyte counts of image cytometer and FCM (cell counts varying from 148 to 2,220/μl) is shown in Figure 9B. A bias of −54 cells/μl and 95% limits of agreement of −261 to 153 cells/μl were found. For CD4 counts higher than 1,500 cells/μl, the FCM gives higher values than the image cytometer, and some results fall outside the 95% limits of agreement. This is probably due to the image analysis algorithm that cannot completely separate all overlapping cells in an image at these numbers.

Next, further tests were performed for determination of CD4 and CD8 count, and CD4/CD8 ratio by analyzing samples from 50 patients. In Figure 10, the regression line and Bland–Altman plot for CD8 count (cell count varying from 18 to 1,242/μl) and CD4/CD8 ratio are shown. A correlation coefficient of 0.96 (P < 0.0001) was found in both cases, and the slope was 0.90 and 1.12 for CD8 count and CD4/CD8 ratio, respectively. For CD4, correlation coefficient was 0.96 (P < 0.0001) and slope 0.98 (graph not shown). Bland–Altman comparison method shows a bias of 33 cells/μl (95% limits of agreement of −106 to 172 cells/ μl) and −0.38 (95% limits of agreement of −1.24 to 0.48) for CD8 count and CD4/CD8 ratio, respectively. This test demonstrates the ability of our instrument to reliably determine both CD4 count and CD4/CD8 ratio, being relevant for staging of HIV infection in infants. In addition, it confirms the good agreement between our instrument and gold standard methods such as FCM in this case. More extensive testing with HIV-negative and -positive patients for CD4 count is reported elsewhere (12).