Flow cytometry (FCM) is the standard technology for characterization and enumeration of cells in a heterogeneous cell mixture. It is routinely used for diagnosis and monitoring of diseases that result in specific changes of a particular cell population. Yet, for characterization of rare populations FCM has its limitations. This explains the emergence of alternative cell analysis technologies for dedicated applications (1–4).
We aim to routinely analyze and enumerate tumor cells that circulate at extremely low concentrations in the blood of cancer patients. We started by using FCM as the analysis platform (5–7). The variable background of the events that classified as tumor cells in a multidimensional gate in FCM (5–7), observed in blood of normal donors urged us to use morphological confirmation. Fluorescence microscopy in combination with immunomagnetic enrichment of circulating tumor cells (CTC) proved to be the right combination. It removes the background and identifies the CTC on basis of their fluorescence signature and morphological features such as round to oval shape, a diameter greater than 4 μm and an intact nucleus that is surrounded by the cytoplasm. This combination formed the basis for CellTracks® Analyzer II. Using this system, the clinical relevance of CTC was demonstrated in prospective multicenter clinical trials. Metastatic breast and colorectal patients with five or more CTC and mestastatic prostrate patients with three or more CTC in 7.5 mL of blood have a significant worse prognosis as compared to patients that have less than 3 or 5 CTC (8–11). Currently, CellSearch® is the only IVD 510(k) system used for serial monitoring of CTC in metastatic breast (MBC), colorectal (MCRC), and prostate (MPC) patients.
The sensitivity of detection and ability to quantify antigen expression of the Analyzer II System is, however, not as good as in FCM. This is caused by trade-offs between spatial resolution, speed of data collection, storage requirements, image analysis, and the wish to keep the total analysis time acceptable for routine applications. Here we describe the development and characterization of CellTracks TDI, an improved image cytometer for potential future use in CTC analysis.
The sensitivity of the CellTracks TDI was tested with 6.8 μm Quantibrite beads (Becton Dickinson, Franklin Lakes, NJ). The four bead populations contained respectively 515, 5,956, 26,653, and 69,045 phycoerythrin (PE) molecules, with CVs of 14.3, 12.1, 14.3, and 13.3%, respectively according to the manufacturer.
The linearity and the variation of the CellTracks TDI were determined using Rainbow Linear Particles, RLP-30-5 (Spherotech, Libertyville, IL). The sample contains a mixture of rainbow particles with five different fluorescent intensities. The intensity drops by factors of 2 from the brightest particle. The particles have a CV of 2.0%. This small CV makes them useful to determine the CV increase introduced by the measurement system. The size of the particles ranges from 3.0 to 3.4 μm.
One tube of Quantibrite beads was reconstituted using 0.5 ml 1× PBS. After reconstitution, 360 μl of the Quantibrite PE sample was transferred to a standard CellTracks cartridge and placed inside a CellTracks Magnest. The assembly was left inverted for 3 h, allowing the beads to sediment to the analysis surface by gravity. This procedure is necessary because these beads are not magnetic and the Magnest only presents magnetic objects at the analysis surface of the cartridge. This resulted in around 50 beads per mm2 that were adhered to the upper surface of the cartridge. Observation under a microscope showed that the majority of the beads remained at the imaging surface, after turning the assembly right side up, even under the influence of gravity. The Quantibrite PE beads could now be measured under conditions similar to a CTC sample.
Rainbow linear particles were diluted in 1× PBS and deposited on a glass slide and sealed with a cover slip. The cover slip was fixed to the glass slide using nail polish to prevent dehydration of the sample.
CTC Positive Samples
Blood samples analyzed using the CellSearch System (Veridex LLC, Raritan, NJ) that contained CTC were reanalyzed using CellTracks TDI. In brief, 7.5 ml of blood was magnetically enriched using ferrofluids targeting the epithelial cell adhesion molecule (EpCAM) and labeled with DAPI, Cytokeratins 8,18,19 PE and CD45-APC using the CellTracks AutoPrep® System (Veridex LLC). The 360 μl enriched sample is then transferred into the cartridge and placed in the Magnest and analyzed on the Analyzer II. The final sample, as prepared by the AutoPrep System, contains ∼85% of the CTC in the original sample together with a small percentage of leukocytes and some debris. These cartridges were then reanalyzed in the CellTracks TDI system.
A schematic representation of the optical layout of the CellTracks TDI image cytometer is given in Figure 1. For excitation three lasers are available that may be used separately or in combination. Emission band pass filters (520/35, 585/40, and 670/30; Semrock, Rochester, NY) are used to eliminate unwanted emission. They are contained in the filter wheel (FW). To merge the beams into one parallel overlapping beam, we use multiple dichroic filters (Semrock). The output of the 16 mW 375 nm solid-state laser (Power Technology, Alexander, AR) is coupled in using a 427 nm dichroic long pass filter (drlp). From the dual line laser (491 and 532 nm, Cobolt AB, Solna, Sweden) only the 20 mW 491 nm output is coupled in using a 503 nm drlp. The 532 nm laserline is filtered out using a 532 nm notch filter (NF/Semrock) because it is not reflected by the triple-band dichroic that is described later and used to direct the laser beams into the objective. The third beam comes from a 30 mW, 639 nm laser (Power Technology, USA). It is reflected by a mirror and passes through the 427 and 503 nm drlp.
The coherence length of the laser lines is much larger than the dimensions of the set-up. The 491 nm line, e.g., has a bandwidth of 30 MHz, hence a coherence length of 3.2 m. To reduce the coherence length, and prevent unwanted interference effects, we use a small motor to rotate the transparent diffuser (12) at 6000 rpm (RD/Suss-MicroOptics, Neuchâtel, Switzerland). It has a pebble-like relief with varying dimensions on its surface, with average pebble dimensions of 40 × 40 μm2. All beams are overlaid at the point where they enter the beam homogenizing optics (Suss-MicroOptics), that create a square homogeneous illumination profile (13–17). The operation of the homogenizer is explained below. After passing the beam homogenizer, the laser light is redirected by a triple-band dichroic filter (tbdr) that reflects the 375, 491 and 639 nm laser lines onto the entrance aperture of the 40×/0.6NA CFI Plan Fluor ELWD infinity corrected microscope objective (Nikon, Melville, NY). These three wavelengths were chosen to be able to excite the DAPI, PE, and APC fluorophores that are used in Veridex's CellSearch® technology. The objective can be moved vertically over a range of 400 μm using a piëzo positioner (MIPOS 500, Piezosystem Jena, Jena, Germany). The objective then focuses the three laser beams on the sample in approximately square illumination spots of 217 × 217 μm2 full width half maximum (FWHM), of which the center 180 × 180 μm2 are used during imaging of the sample. The resulting (maximum) irradiance at the object plane is 10.8 W cm−2 at 375 nm; 24.0 W cm−2 at 491 nm, and 33.6 W cm−2 at 639 nm.
Part of the emitted fluorescence is collected by the objective. It passes through the triple band dichroic filter. A motorized filter wheel (Thorlabs, Newton, NJ) selects the correct emission filter for each particular fluorescent probe. The emission light is focused by a 160-mm achromatic tube-lens (TL/Linos Photonics, Goëttingen, Germany) onto the high sensitivity 12-bit Peltier cooled ORCA C4742-95-12ERT TDI camera (Hamamatsu, Hamamatsu City, Japan) with 1344 × 1024 pixels of 6.45 × 6.45 μm2. This camera can both operate in TDI and frame transfer mode. A blue 470-nm LED (Philips-Lumileds, San Jose, CA) underneath the sample is used for bright-field illumination.
When scanning the sample at a typical speed of 1 mm s−1, each pixel in the object space (0.2 μm) is illuminated during 200 μs. The laser beam has a diameter of 700 μm and hits the diffuser at 10 mm from its center. As the diffuser rotates at 6,000 rpm, the beam travels in 200 μs over 1256 μm of the diffuser surface, which is 1.8 times the beam diameter and 31 times the average pebble size on the diffuser. This creates sufficient randomization of the phase of the beam during the time a pixel is illuminated and thereby greatly reduces interference effects due to coherence of the laser beam.
The beam homogenizer uses two square micro-lens arrays that consist of a periodic structure of micro-lenses, with pitch PMLA (Fig. 2A). The first array (MLA1) focuses the beam on the second array (MLA2), resulting in multiple point sources. The light from all these separate sources is collected by the microscope objective and overlaid in the focal plane where they form a quadratic illumination profile with a homogeneous intensity (flat top profile) with a size given by (18):
where, PMLA = 0.50 mm, fobj = 5 mm, the focal length of the microscope objective, f1 = f2 = 15.54 mm, the focal lengths of the micro-lens arrays and a12 = 15.54 mm, the distance between the two micro-lens arrays.
When all components are positioned according to the above described distances, the resulting illumination profile is approximately square with a FWHM of 161 μm. However, the actual width of the illumination profile, with a CV of <5%, is then about 130 μm. This results in the need to scan 21 “strips” to cover an entire cartridge. To reduce this number and still be able to see the edges of the illumination profile for focus determination purposes, a12 may be slightly increased. To avoid large aspect ratios of the images and reduce the number of events that are located at the border of an image, the images are stitched together in groups of four. The accuracy of overlay of fluorescent channels is limited by the stage accuracy, which is 0.2 μm.
Feed Forward Focusing
The analysis surface is never perfectly flat, resulting in a variation in focus position in the Z direction when scanning the surface. Prior to a scan, the focus positions are determined automatically at a grid of 6 by 5 positions on the surface. An automatic focusing algorithm uses the reflection of the red laser profile from the glass-sample interface. It integrates the reflection over the 400-μm range of the piëzo positioner using the TDI camera. This generates an image that contains the profile of the illumination spot for every part of the 400-μm range. A custom made algorithm is then used to find the position were the illumination profile has the sharpest defined edges, indicating the focus position. These positions are then fitted by a 3D spline that estimates the correct focus position with an accuracy of 0.5 μm (smaller than the depth of field of the microscope objective, which is 1.5 μm for light with a wavelength of 550 nm) at each point on the surface. During a scan, the 3D spline fit is used to focus the objective with the MIPOS operating in feed forward mode. The bandwidth of the piëzo system was determined by measuring the response at increasing frequencies. The −3 dB point was found to be at 4.5 Hz, resulting in a “bandwidth” of 1.8 mm s−1. At a default scan speed of 1 mm s−1, this translates to 1.8 mm mm−1, which is more than enough to follow the variations in height of a typical cartridge, which are in the order of 5 μm mm−1.
The cartridge is scanned through the 180 × 180 μm2 illumination profile, see Figure 3A. It is moved in the X and Y direction by two stacked servo stages, featuring brushless DC motors, (M-605.2DD from Physik Instrumente, Karlsruhe, Germany) with a maximum continuous scan speed of 50 mm s−1 and a travel range of 50 mm, resulting in a total area of 2,500 mm2 that can be covered. The complete analysis surface is scanned in the Y direction in multiple adjacent strips with a width of 180 μm. At the end of each strip, the cartridge is moved 180 μm in the X direction and to the start of a new line in the Y direction. Then, the next strip is scanned. Precautions are taken to prevent effects due to backlash of the scan system. Each strip scan starts by first moving the slide to a fixed position where the beam is outside the area to be scanned. Next, the scan is started and data are recorded not sooner than after passing a second fixed position on the surface. This procedure is followed to ensure that the stage has achieved its final speed when data are recorded. The overhead time introduced by the acceleration and deceleration of the stages is about 1 s strip−1. Two built-in encoders with a resolution of 0.1 μm continuously determine the actual position of the stages.
The CCD camera operates in TDI mode, resulting in a continuous readout of collected charges. This mode is illustrated in Figure 3B. When an event is scanned through the illumination profile, the CCD pixels are continuously collecting charges. The collected charges of each pixel row are transferred in parallel one row down to the next pixel row when an external line trigger is received. This trigger is derived from the encoder that reports the actual stage position. The parallel charge shift rate in the CCD is synchronized with the continuous motion of the stages. When the transferred charges reach the last row of CCD pixels, the total collected charge for each pixel is moved to the amplifier and A/D converter. The charge of each separate pixel stems from the integrated light intensity emitted by one corresponding pixel in the object space while the event moves through the illumination profile. As the object moves, the accumulated charges move in the corresponding direction on the CCD, with the same average relative velocity. This makes light integration and CCD readout a continuous process preventing dead time. The resulting image of a strip is 900 pixels wide. Its length is only limited by available memory on the frame grabber, which is 16 MB. To obtain workable images for analysis in ImageJ and to fit inside the frame grabber buffer, we limit the number of pixels that is read-out into one image in the scan direction to 10,000, amounting to a total size of 12.9 MB. Each image is then transferred to a larger reserved buffer in RAM memory, from where it is written to disk.
Discrete Final Magnification
To prevent blurring and obtain sharp images, precise alignment of the direction of stage movement and the direction of the columns on the CCD is essential. Also, the average speed of the moving object should be matched with the line transfer rate of the camera. Although it is possible to synchronize the camera with a function generator that matches the average scan speed, optimal resolution is only obtained when applying the encoder signal of the translation stages. Small variations in scan speed can then be synchronized with the camera line transfer rate for each position. As the encoder has a resolution of 0.1 μm, the choice for the magnification, M, is then limited to discrete values given by:
where δp is the size of the pixels on the camera and δe the encoder resolution. With the use of a counter the required object pixel resolution can be obtained as a multiple of the encoder resolution. In our setup a sampling of δs = 0.2 μm pixel−1 is used requiring a magnification of 32.25×.
To operate the CellTracks TDI, all different parts have to function in the proper manner at the correct time. To this end, a dedicated software package was written in LabView (National Instruments Corporation, Austin, TX), a graphically oriented programming language. LabView is very appropriate for control of the cytometer as it has many mathematical subroutines and supports an extensive amount of instrument drivers for hardware control. The program can easily be adjusted to accommodate new hardware components. Flowcharts of the dedicated software program are given elsewhere (19).
A scan of a cartridge sized imaging surface (29.7 × 2.7 mm2) at 0.2 μm lateral resolution in three fluorescence and one bright field channel generates 16 GB of raw 12-bit image data. This data needs to be processed efficiently to remove the >99% of imaged surface area that doesn't contain an event. To this end, we developed a dedicated image analysis routine using ImageJ, a public domain Java program, originally developed at NIH (20, 21).
The routine corrects for inhomogeneous illumination, removes the background and selects the events of interest from the raw data. The selected events are then combined and further analyzed to obtain quantitative parameters.
The image analysis procedure is divided in three distinct steps. First, the raw 900 × 10,000 pixel images that have been recorded during a scan are corrected for the variations in the normalized (to the average intensity) illumination profile (22–24) in the direction perpendicular to the scan direction. No correction is needed in the scan direction as the TDI camera averages these intensity variations. Then the background is subtracted by applying the rolling ball method (25) using a diameter of 80 μm. This diameter is about three times larger than the maximum diameter of the future positive events (CTC). This was chosen to ensure that no signal is removed during this step, even when two large events (30 μm) are lying side by side.
Second, either the fluorescence or the bright field image is used to select events of interest, defined as those events that have a pixel intensity above an appropriately chosen fixed threshold, e.g., 50. The fluorescence- and bright-field image data of all events of interest are then combined in one multipage tiff file. This second step drastically reduces the amount of data that needs to be saved, typically by a factor of 100.
As shown in Figure 4, for a Quantibrite bead, some pixels in the background may have an intensity above a threshold level of three times σp (average pixel noise) and some pixels within the bead have an intensity below this threshold. Using this criterion for threshold determination, results in incomplete selection of the bead. This can be prevented by applying a Gaussian blur filter that averages the pixel intensity within a specified radius for each pixel. However, it reduces spatial resolution, but averages out the noise thus enabling to segment the particles at a lower threshold limit and obtain a proper mask for the event. When a bright field image of an event is used to create a mask, the detection limit is independent of the fluorescence level of the event in a different channel. This enables analysis of events with a very dim fluorescence signal.
Third, the last image analysis step creates a mask that defines, for each separate event, its boundaries in all channels. To this end, the fluorescence images of all events are thresholded by the histogram-based Otsu algorithm (26). It determines a threshold by maximizing the between-class variance of pixel intensity in an assumed bimodal histogram and is relatively fast because only the histogram is needed. In the resulting binary image the holes are filled. Then it is used as a mask for the original image, to define the edges.
When a mask has been defined for each image of each event, quantitative and morphological parameters, such as size, total fluorescence intensity, and circularity are calculated. These quantitative parameters are presented in dot plots where each event is linked to the actual image. This enables review of the sample by means of quantitative and morphological parameters. If desired, each particular set of parameter values recorded can directly be compared to the image of the event from which they were determined.
The size of the flat top intensity profile created by the beam homogenizer was adjusted by slightly varying the distance between the two microlens arrays (a12 as indicated in Fig. 2A). Care was taken not to overfill the aperture of the lenses in the second microlens array. Overfilling results in crosstalk between the lenslets and a less sharply defined border of the flat top profile. To determine the size and the intensity profile, we used a homogeneous layer of acridine orange. The resulting illumination profile, shown in Figure 2B, then has a flat top with edges that are not as steep as theoretically possible, but now with a width of 180 μm that can be used for imaging. During scanning, only the indicated area with a size of 180 × 180 μm2 is captured from the CCD camera. The intensity outside this area drops rapidly, which is important to prevent bleaching of adjacent areas when scanning the sample. If needed, the size of the illumination area can be adjusted further by choosing a different objective.
A difference in motion between the TDI CCD and the stage will result in an increase of the effective sampling aperture L and leads to blurring of the image in the Y direction (27). The angular misalignment between the CCD position and the stage scan direction affects the resolution perpendicular to the scan direction. A small angular misalignment θ results in a pixel overlap
where NTDI is the number of TDI stages (1,024) and δp is the pixel size in the scanning direction. By adjusting the CCD and the stage scan direction with accuracy of 0.03° or better this blurring effect could be limited to 0.5–1 pixels.
During stage motion the stage position appears to be moving with a slightly oscillating velocity V that has a standard deviation of 0.8 μm around the commanded value. This creates a blurring pixel overlap, given by , equal to 8 pixels in case of a fixed line rate. For this reason, an encoder is used to measure the stage position continuously. The encoder triggers the camera each time when the stage position has moved by 0.2 μm and thus ensures the synchronization between stage motion and CCD charge transfer. To obtain a 1:1 relation between stage movement and CCD charge transfer, the magnification was adjusted to 32.25× as accurately as possible.
The residual mismatch in magnification then is pixels. This mismatch appeared to result in a blurring effect of 1 pixel.
Another cause of blurring that cannot be prevented is due to the discrete movement of the charges of the TDI-CCD while the object moves continuously. This also results in a blurring of 1 pixel in the TDI scan direction.
To determine the total degradation of the resolution due to all these factors, the modular transfer function (28–31) (MTF) or spatial frequency response (SFR) was measured for the entire optical system and the TDI scanning system. From the MTF curves we can derive the actual resolution of the complete optical system. According to the Rayleigh criterion, the resolution is the inverse of the spatial frequency for which the MTF drops below 9%.
A cartridge with a slanted step edge was filled with an Acridine Orange solution. The upper surface was lined with 15-μm wide nickel lines with a spacing of 30 μm to partly mask the fluorescence of the Acridine Orange. The slanted step edge was imaged under an angle of ∼5° parallel and perpendicular to the scan direction to determine the MTF in both directions. To image the Acridine Orange layer, the green laser line was used for excitation in combination with a 580DF30 band pass filter. The images were analyzed with Matlab and a Slant Edge Analysis Tool, sfrmat 2.0 (32, 33). The resulting MTF curves of the optical system in the scan (Y) direction, at a wavelength of 580 nm, are shown in Figure 5. They illustrate the difference in image degradation between the camera operating in frame transfer mode (2), TDI mode with encoder triggering (3), TDI mode with fixed line rate triggering (4), and the theoretical limit of the MTF (1). When scanning in TDI mode with encoder triggering the system has a slightly lower MTF in the scan direction as compared to acquiring images in frame transfer mode. This degradation is due to the pixel blurring and results in an ∼10% decrease in resolution. Scanning at a fixed line rate of 5 kHz at an average speed of 1 mm s−1 shows severe degradation and results in about 60% decrease in resolution. This is primarily caused by a mismatch between stage speed and camera TDI row shifting. The MTF perpendicular to the scan direction is almost equal for all camera modes.
In our case the optical resolution, when scanning using TDI mode combined with encoder triggering, at 1 mm s−1, is 0.72 μm in the X direction and 0.76 μm in the scan direction (Y) for light with a wavelength of 580 nm. If a fixed line rate is used, the optical resolution increases to 1.5 μm in the scan direction, while the resolution in the X direction stays the same. When using the frame transfer mode of the camera, the resolution in the scan direction improved to 0.66 μm while the resolution in the X direction increases slightly to 0.69 μm. Finally, the ideal optical system would have a resolution in both directions equal to the wavelength of the light that is used.
Acquisition and Analysis Time
The total acquisition time in the CellTracks TDI is mainly determined by the scan speed. Overhead time due to stage repositioning is kept to a minimum by the use of the TDI camera. A default cartridge is scanned for four colors; DAPI, PE, APC, and Bright field. For each color, 16 strips, 30 mm in length, are scanned to cover the entire analysis surface. At a default scan speed of 1 mm s−1, each color takes 8 min in pure scan time. Some overhead time is introduced by moving the sample to the start of a new strip and by the time it takes the stages to reach a constant velocity before scanning can begin. On average this overhead time is 1.25 min, leading to a total scan time including overhead for one color of 9.25 min. For four colors, the total scan time then amounts to 37 min. Without TDI camera, using a frame transfer camera, the total scan time for four colors would be 80 min.
Before a sample can be scanned, the borders of the cartridge and the feed forward autofocus settings have to be determined. This takes around 3 min, leading to a total acquisition time of around 40 min for one cartridge in four colors.
After the acquisition, the data is processed and analyzed as explained before. It takes 1 h and 40 min to process the vast amount of data (16 GB) obtained by scanning a cartridge, with a total surface area of 81 mm2, at a sampling resolution of 0.2 μm pixel−1. The analysis computer is equipped with a Core2 Duo processor at 2.1 gHz, 4gB RAM and Windows XP.
The fluorescence sensitivity of the CellTracks TDI for PE was measured using Quantibrite PE beads (34). The beads were excited at 491 nm and a 580DF30 emission filter was used. Figure 6A, shows histograms of the mean PE signal intensity as measured at a speed of 1 mm s−1, without neutral density (ND) filter and with a ND0.5 filter. The values were calculated by subtracting the background-, dark current-, and bias contributions from the total signal of the object and dividing by the area of the bead (1,020 pixels). All four bead populations were resolved from the background. The CV is 12–16% for the four peaks and is primarily determined by a variation of the number of PE molecules/bead.
Measurements of the Quantibrite PE beads with a flow cytometer (FACSAria II, BD, San Jose, USA) show variations of about 13% and are thus of the same order as obtained with the CellTracks TDI. Panel B of Figure 6 shows the relation between the mean PE signal intensity and the known number of PE molecules. Clearly, a linear correlation exists confirming the linearity of our system. The lower limits of detection are given by the position where three times the standard deviation in the background coincides with the extrapolated lines. The detection limit for PE molecules on a bead with a diameter of 6.8 μm at a scan speed of 1.0 mm s−1 is ∼120 without ND filter and ∼340 when a ND0.5 filter is used.
The linearity and the variation of the CellTracks TDI were measured with Rainbow linear particles scanned at a scan speed of 1 mm s−1. The 491 nm green laser line was used for excitation after attenuating the beam by a ND0.4 filter to prevent saturation of the brightest peak. Under these conditions the peak pixel intensity has a mean value of ∼1,200 which is ∼1/3 of the full camera range. The recorded histogram is shown in Panel C.
The average CV of the peaks is 4.6% with a minimum of 4.4% for Peaks 1 and 3 and a maximum for Peak 4 of 5.1%. The system has a linear response and a linear regression through the plot values results in R2 = 0.9998 as shown in Panel D. The ratio between the bead populations is ∼2.0 as specified by the manufacturer.
The combination of immunofluorescence quantification and morphologic examination of CTC candidates is illustrated in Figure 7. Panel A shows a dotplot of total CD45-APC intensity versus total Cytokeratin-PE intensity of events passing a threshold on the PE intensity and Panel B shows the same dotplot now with a threshold on the DAPI intensity (not shown). These quantitative values are determined by a region of interest around each event calculated by an Otsu threshold algorithm. The total intensity values are calculated by summing the intensities of all pixels inside the region of interest. The larger number of events in Panel B is mainly caused by the leukocytes staining with both DAPI and CD45-APC, but lacking Cytokeratin-PE. The larger number of events in the region staining with Cytokeratin-PE and relatively low CD45-APC staining in Panel A as compared to Panel B is caused by the omission of a DAPI threshold. Eight events are highlighted in Panels A and B and the DAPI, Cytokeratin-PE, CD45-APC, bright field and overlay images associated with these events are shown in Panel C. Image 1 and 2 show morphological features that classify the events as CTC by the CellSearch definition (35), the cell in Image 2 however shows some staining with CD45-APC, but the staining pattern is not that of a typical leukocyte. Image 3 shows an apoptotic CTC, as illustrated by the punctuated PE staining and dimmer DAPI staining than the cells in 1 and 2. Image 4 represents a tumor micro particle (36, 37), and Image 5 resembles a leukocyte with nonspecific PE staining. Image 6 appears as debris, Image 7 shows morphological features associated with a granulocyte and Image 8 has morphological features associated with a lymphocyte.
Cartridges of four patients were analyzed on the CellTracks TDI and the results are shown in Figure 8. Events that were classified as CTC by an expert reviewer, according to the CellSearch definition, are indicated as red dots and are surrounded by events that were not classified as CTC. The CellSearch definition contains morphological and size criteria, e.g., round to oval morphology and minimum size of the cell. The four dotplots illustrate that CTC do not appear in a confined region in a CD45-APC versus Cytokeratin-PE dotplot. This illustrates the importance of having high resolution fluorescence- and bright field images of events in a patient sample for visual confirmation and classification by an expert reviewer to obtain a reliable CTC enumeration.
A fully automated image cytometer was developed for routine analysis of samples in which only very few positive events are expected amidst debris and leukocytes. In this study, the system was optimized for the detection of immunomagnetically enriched CTC from blood of cancer patients. For this application the definition of a CTC is of importance making it imperative to avoid false positive events or miss true positive events. This means that for identification we have to use both morphological- and immunofluorescence parameters, determined at relatively high resolution and sensitivity. The time needed for analysis should be within limits that are reasonable for routine application in a clinical chemistry setting. Therefore, in designing the cytometer a compromise had to be made between resolution, analysis time, sensitivity, and number of measured parameters. Also the overhead time in scanning and analyzing had to be minimized. The image cytometer has a proven scan resolution of 0.76 μm, for light with a wavelength of 580 nm, and images three different fluorescent- and one brightfield channel for a cartridge. It scans a surface of 2.7 × 30 mm2 in 9.25 min for each parameter amounting to 37 min for DAPI, PE, APC and brightfield combined. The sensitivity of the optical system is 120 PE molecules at a default scan speed of 1 mm s−1.
These results are obtained by using three laser lines, a 40×/0.6NA objective and a beam homogenizer to create a high intensity square homogeneous illumination profile for fluorescence excitation of a wide range of fluorophores. Different laser lines can be added to the system, provided that a suitable multiple band dichroic filter can be produced. A CCD camera operating in TDI mode is used to lower the overhead time in scanning and to increase the sensitivity. TDI scanning is performed in a continuous motion, in synchronization with the scanning of the stages. Microscope based systems that use lasers, such as confocal microscopes and laser scanning cytometers, use a scanning mirror for illumination. The advantage of using a beam homogenizer, as in our system, is the simultaneous homogeneous illumination of the entire field of view. The combination of laser illumination and beam homogenization allows for control over the laser power in a well defined area, thereby avoiding bleaching of part of the sample that is not currently being imaged. Using a laser with, e.g., double the output power would also double the sample illumination, something which is not possible in mercury arc lamp systems. This would reduce the integration time needed for the dyes that are used in the CellTracks TDI by a factor of 2 as these are well away from being limited by bleaching (19). Although mercury arc lamps are frequently used as a light source in fluorescence microscopes, the advantage of having multiple excitation wavelengths is overshadowed by the limited lifetime of the bulb, the need for realignment and the inhomogeneous illumination. The latter could be overcome by the use of a beam homogenizer.
A high power LED provides the illumination for bright field imaging in the image cytometer. However, image quality in the bright field images is not optimal as can be seen in Figure 7C. A barcode-like pattern is visible in these images. This is due to the unbound magnetic particles that are present in the sample, which are also moved to the analysis surface by the magnetic field. There they form these characteristic patterns that obstruct clear brightfield imaging. We are currently developing methods to remove the unbound magnetic particles from the samples to improve both brightfield and fluorescence imaging.
The image cytometer allows us to detect immunomagnetically enriched and fluorescently labeled CTC quickly and accurately from blood samples of patients. Quantitative measurement of fluorescence in combination with morphological examination permits for a better discrimination within the CTC candidates. Thus, it creates new opportunities for a more accurate determination of the clinical meaning of CTC in blood of cancer patients. For rare event detection, flow cytometry-based systems do not analyze the complete volume of the sample. This shortcoming is addressed in the image cytometer by the use of an analysis cartridge in which all the magnetically-labeled cells in the sample are moved to the analysis surface. Moreover, the analysis cartridge provides the opportunity to revisit cells after initial analysis. The system can automatically relocate events of interest and perform further detailed analysis, even under different sample conditions. It can create Z-stacks of images of selected events to perform 3D analysis and record time-lapse series. This can be used to extract more information from the analyzed cells such as the position of intracellular structures or FISH probes (38) which for example cannot be achieved by flow cytometry-based systems.
Future improvements of the cytometer will include reduction of image acquisition time of multiple fluorophores by parallel imaging. This requires the addition of one or more TDI cameras and a beam splitting element to the setup. We are currently in the process of adding a second TDI camera to the system to reduce imaging times by a factor of 2. Alternatively, a TDI camera may be used in which different areas on the CCD can be read out separately such as used in the ImageStream system (39). A different option is to use a TDI camera operating at a higher line rate. The maximum line rate of the current TDI camera is 8.8 kHz, resulting in a theoretical maximum scan speed of 1.76 mm s−1. A newer TDI camera, now available from the same manufacturer, features a maximum line rate of 31 kHz, raising the theoretical maximum scan speed to 6.2 mm s−1. This is almost a factor of 4 higher than the current maximum scan speed. The latter option would also allow the use of higher laser power for illumination, which is not possible now because of the saturation of pixels at higher intensities. If desired, a fourth fluorescence parameter can be added. This requires the use of a quad-band dichroic filter. A drawback is that such filters often have lower efficiency for the separate dyes as compared to a triple band dichroic. Apart from the mentioned hardware changes, the image analysis can be made faster by using a separate dedicated analysis computer with multiple processors at high clock speed.
We estimate that the overall reduction in time needed for analyzing a cartridge of 2.7 × 30 mm2 using three fluorescence- and one brightfield image could be a factor of 8, bringing the time needed to analyze one sample well below 5 min.
Even though we designed and built the image cytometer in the first place to improve the method to determine CTC in blood samples, it is obvious that many applications of the instrument exist. In fact, the cytometer can be adapted easily for the analysis of other cells where in a routine setting high resolution imaging and quantitative multispectral fluorescence characterization of cellular samples need to be combined.