SEARCH

SEARCH BY CITATION

Keywords:

  • image cytometry;
  • slide-based cytometry;
  • whole slide imaging;
  • virtual microscopy;
  • digital slide

Abstract

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

Slide-based image cytometry (SBC) has several advantages over flow cytometry but it is not widely used because of its low throughput, complicated workflow, and high price. Fully automated microscopes became affordable with the advent of whole slide imaging (WSI) and they can be transformed into a cytometer. A MIRAX MIDI automated whole slide imager was used with metal-halide and light emitting diode (LED)-based fluorescent illumination, filter block changer, and a cooled monochrome charge coupled device camera. The MIRAX control software was further developed for fluorescent sample detection, autofocusing, multichannel digitization, and signal correction due to nonuniform illumination. Fluorescent calibration beads were used to verify the linearity of the system. The HistoQuant software package of the MIRAX viewer was used for image segmentation and quantitative analysis. The data was displayed by the histogram, scatter plot, and gallery functions of the same program. Fluorescent samples can be reliably detected, focused, and scanned. The measured integrated fluorescence showed linearity with exposure time and staining intensity. Automated fluorescent WSI with stable LED illumination and high-quality homogeneous fluorescent slides can be used conveniently for SBC. © 2009 International Society for Advancement of Cytometry


INTRODUCTION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

Although the first flow cytometry (FCM) and image cytometry experiments were made approximately in the same time in the 1950s (1, 2), FCM became more widely adopted. Slide-based image cytometry (SBC) has several advantages over FCM. It provides morphological information from tissue structure to subcellular components. Based on the measured data, selected cells can be re-examined with different imaging techniques. Samples can be re-stained and re-scanned to acquire more parameters and the analysis of cell clusters and tissue is also possible (3–5). Despite the advantages of SBC fluorescence labeled large cell populations are analyzed today dominantly by FCM for the following reasons. FCM has a throughput of 10,000 to 100,000 cells per second in comparison with SBC's 100–1,000 (3, 6) cells per second. In SBC measurements specimen position, outline and focus on the microscope slide has to be determined precisely, which slows the workflow even further.

The development of computer and charge coupled device (CCD) camera technology about 20 years ago made it already possible to have imaging cytometer systems with acceptable throughput and data handling capability (7–11). Since that time different systems were developed to push the boundaries of image cytometry. The slide-based systems include laser scanning cytometry (3, 12–15), standard (16–19), enhanced wide-field (6, 20–23) and confocal fluorescent microscopes (4, 5, 24). FCM can also be extended with imaging capabilities (25–27).

A laser scanning cytometer (LSC) uses lasers to scan a slide. This solution provides good results because the field of view (FOV) is illuminated evenly and the resolution is high. This technology developed by Compucyte (Cambridge, MA) is not confocal on purpose to collect fluorescence from the full depth of field. LSC has medium throughput and it is expensive due to the lasers and photomultipliers used in the system.

Q3DM (San Diego, CA) enhanced a wide-field fluorescent microscope with phase-contrast based focusing together with high speed electronics and stabilized mercury arc lamp to increase imaging speed and quality (6, 20, 21). The core group of former Q3DM employees developed recently a differential interference contrast autofocus system (22) for fluorescence microscopy and a volume camera based fast autofocus system for a continuous-scanning automated microscope (23).

Ecker et al. (4, 5, 24) developed a confocal microscope-based tissue cytometer. Tissue samples can be better segmented using the confocal technique because it images only a thin layer of the specimen in contrast with LSC and the cells do not overlap on the recorded images. On the other hand, fluorescence is not collected from the full depth of field therefore precision and throughput is low and confocal microscopes are the most expensive modalities for image cytometry.

The most affordable solution is a standard fluorescent microscope with appropriate software. Based on our earlier work (17, 18), Carl Zeiss (Carl Zeiss MicroImaging Gmbh, Jena, Germany) released recently the AxioVision scanning fluorescent microscope application. Recently, Hennig et al. described a microfluidic chip scanning method based on a similar configuration (19). Regular fluorescent microscopes are limited to work with one slide and their FOV is illuminated nonuniformly unlike the laser-based systems. Illumination nonuniformity can be compensated but the complete measurement workflow requires several steps of human interaction for slide loading and unloading, compensation image recording, scan area designation, and main focus setting.

In recent years, another imaging microscopy field developed very rapidly. Brightfield virtual microscopy or whole slide imaging (WSI)—the naming convention is not consistent yet—became commercially available and starts to be established in pathology (28–32). WSI means that complete sections, cytospins, and smears can be digitized automatically in high-enough resolution that is appropriate for diagnosis. There are three systems on the market currently that are capable of fluorescent WSI, the MIRAX SCAN and MIDI from 3DHISTECH (3DHISTECH, Budapest, Hungary) and Carl Zeiss and the Hamamatsu Nanozoomer (Hamamatsu Photonics K.K., Japan). The MIRAX system developed by our group utilizes area cameras and is capable of digitizing a sample on nine fluorescence channels. The NanoZoomer uses a time-delay-and-integration CCD camera (TDI) and is capable of scanning on three fluorescence channels.

The goal of this study is to describe the working principle of the MIRAX fluorescent whole slide imager and to apply it for image cytometry.

MATERIALS AND METHODS

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

Samples

For testing and calibration of the system, 10 μm diameter cytometric calibration beads were used (Immuno-Brite Fluorospheres, Art. No. 6603473; Beckman Coulter, Fullerton, CA). The kit includes five different populations with 100%, 25%, 6.25%, 1.5625%, and 0% intensities. For evaluation of clinical samples, residual samples from young, cancer-free patients were used. Mononuclear cells were isolated from EDTA anticoagulated blood with standard density gradient centrifugation (Histopaque-1077; Sigma-Aldrich, St. Louis, MO). After removal of mononuclear cells from the layer floating on the Histopaque, the cells were washed three times by PBS and counted in a hemocytometer (Buerker chamber). A total of 10 μl of the cell suspension was smeared on a microscope slide and dried. Cells were stained in 100 nM Hoechst 33258 for 20 min, and washed in PBS.

ProLong Antifade (Molecular Probes, Eugene, OR) was used on dry smears as recommended by the manufacturer to minimize fluorescent fading and to permit multiple scanning.

Hardware

We used a MIRAX MIDI automated digital microscope shown in Figure 1. The MIRAX MIDI can scan in brightfield and in fluorescence with our development. It has a slide loader mechanism for 12 slides and a 200 μm focus range. The system had a Zeiss 20× Plan-Apochromat, NA 0.8 dry objective. Three high efficiency Carl Zeiss fluorescence filter cubes were used: DAPI (Filter Set 49), FITC (Filter Set 38HE), and Rhodamine (Filter set 43HE). For the quantitative fluorescence measurements, a Carl Zeiss Colibri light emitting diode (LED) light source was used with 365 and 470 nm LED modules. For general scanning tests, a Carl Zeiss HXP 120 metal-halide short arc lamp was used that can be fiber coupled to the Colibri lamp. The illumination pathway was used from the Zeiss AxioScope 40 and the microscope had a 10 position filter wheel. For image capture, a Zeiss AxioCam MRm Rev.3 monochrome CCD camera was used. The camera has 1388× 1040 pixels and 6.45 μm × 6.45 μm pixel size, 12 bit digitization, 17,000 electrons full well capacity, and single-stage Peltier-cooling. With the 20× objective, one pixel imaged 0.3225 μm × 0.3225 μm area of the specimen. A DFK F2104 camera with 640× 480 pixels was used (The Imaging Source Europe GmbH, Bremen, Germany) for slide preview capture.

thumbnail image

Figure 1. MIRAX MIDI and its hardware components. (A) The MIRAX MIDI system with the AxioCam MRm mounted on the top and a slide tray inserted on the right. Left to the system, HXP-120 fluorescent light source is shown. (B) Components inside the microscope: A, preview camera; B, LED-based slide background illumination for preview camera; C, LED-based illumination for label area; D, objective; E, fluorescent illumination unit; F, slide holder and slide; G, slide holder tray; H, slide loader; I, slide loader stepper motor; J, Y lead screw and rails of slide stage; K, halogen lamp-based transmitted light illumination unit. (C) The schematic light path of the system. The same letters are used as in (B) where applicable. L, mirror; M, condenser; N, optics for transmitted Köhler illumination; O, halogen light source for transmitted illumination; P, fluorescent filter cube; Q, additional optics for epifluorescent illumination; R, fluorescent light source; S, tube lens; T, AxioCam MRm. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

Download figure to PowerPoint

The software ran on a PC system with dual Intel Xeon 2.8 GHz processors, 2 GB RAM, and 500 GB hard drive. The operating system was Microsoft Windows XP (Microsoft Corp., Redmond, WA).

Software Development Tools

The software is an extension of the MIRAX SCAN control program developed in Microsoft Visual Studio 6 and Borland C++ Builder (Borland Software Corp., Scotts Valley, CA). We used the MIRAX viewer digital slide viewer for displaying the slides.

Algorithms

To digitize fluorescence-labeled samples on a slide its location, focus position, and exposure time in every channel has to be determined.

Sample Detection, Localization, and Mapping

The MIRAX MIDI is equipped with a preview camera to grab a low-resolution image from the slide to determine areas for imaging. The preview camera has its own objective and separate optical path from the main camera and separate optical path from the main fluorescent illumination. In the preview position, the sample is illuminated from the back by an LED panel (Figures 1B and 1C). Fluorescent samples have low contrast and are not detectable directly by the preview camera. To overcome this limitation, the fluorescent imaging software requires a continuous line marking around the sample on the slide with a black marker pen. In the process of scanning, the system first moves the slide in the optical path of the preview camera. It detects the marker pen and calculates which FOVs have to be digitized by the high-resolution main camera.

A grid is placed virtually on the sample before imaging the sample. The sample is focused and exposure times are measured in every channel on this grid. The focus values are interpolated to determine every FOV's own focus level. In every channel, the shortest grid point exposure time is selected for scanning. The distance of the grid points can be set manually in units of FOVs. A typical value is 3 FOV.

Mapping time is shortened by limiting the full 200 μm focus range to 50 μm. The algorithm goes through the grid points from the center of the scan area following a spiral path. The focus level of the first field of view that contains sample will be the middle of the limited range. A FOV is considered to contain sample if there are values above column 50 in the pixel value difference histogram. The generation of the histogram is described in the Sharpness calculation section.

Focusing

The MIRAX MIDI has a 200 μm focus range. In fluorescent microscopy, both excitation and emitted lights are focused by the objective. As the fluorescent sample gets out of focus, its brightness decreases as shown in Figure 2. To adapt to the varying light intensity, the exposure time is continuously adjusted during focusing. An intensity range is defined and the algorithm keeps the brightest pixels always in this range. The bottom of the range has half the brightness as the top. Three different ranges can be selected in the software. Range one is from 32 to 63 which is the default, range two is from 64 to 127 and range three is from 128 to 255. The autofocus algorithm first goes through the whole range in 4 μm steps then fine focuses around the best position. Figure 3 shows the flow chart of the coarse focus integrated with the continuous exposure change.

thumbnail image

Figure 2. Exposure time as a function of focus. In fluorescent microscopy, the exposure time strongly increases as the sample gets out of focus. To demonstrate this effect, the exposure time was calculated for a field of view with a single bead at different focus levels throughout the focus range. At 76 μm, the exposure time was 9 ms and at 164 μm 732 ms.

Download figure to PowerPoint

thumbnail image

Figure 3. Coarse focus flow chart. The flow chart shows how the autofocus algorithm continuously adapts the exposure time as it goes through the focus range during coarse focusing.

Download figure to PowerPoint

The focus algorithm selects the best focus position based on the sharpness value of the image. A higher sharpness value means a better focused image. Sharpness calculation is based on pixel value differences. To lower noise and cancel small artifacts, the image is shrunk by averaging 3.33 × 3.33 pixels. After shrinking, the algorithm goes through the image and calculates the difference between every pixel and its neighbor to the right. The fifth power of every difference is calculated and added to the sharpness value of the image. If, for example, the difference of two pixels was 10 then 105 is added to the sharpness value.

From two images with different exposure times but equivalent sharpness value the one with shorter exposure is the better focused. Taking this in to account, the calculated sharpness values are divided by the exposure time.

The reliability of the focusing algorithm was tested by autofocusing 20–20 FOVs of five different tissue samples using the Rhodamine filter. After autofocusing, the FOVs were fine focused again manually in 0.2 μm steps and the difference between the manually and automatically determined focus values was calculated. Table 1 summarizes the results.

Table 1. Focus level differences
Difference in micrometer (μm)Number of field of views
  1. The left column shows the difference in micrometer (μm) between the automatically found focus levels and the manually fine focused position. The right column shows how many field of views were found with the given difference.

−0.83
−0.417
054
+0.422
+0.84

Image Compensation

For the correction of the illumination nonuniformity, a special compensation slide was prepared by the Fraunhofer Institute for Applied Optics and Precision Engineering IOF (Jena, Germany). A 1.45 μm thick polymethylmethacrylate layer on a glass slide contained the following laser dyes: [Radiant Dyes Laser & Accessories GmbH, Wermelskirchen, Germany; article number, name, concentration (mol/l)]: 044, Coumarin 2 (C450), 3 × 10−3; 072, Coumarin 545, 1 × 10−4; 084, Rhodamine 6G (Rh590), 1 × 10−4; 087, Rhodamine 101 (Rhod640), 1 × 10−4; 102, Oxazin 4 (LD690 Perchl.), 1 × 10−4; 101, Nile Blue Perchl., 2 × 10−5; 119, Rhodamine 800 (LD800), 2 × 10−3.

The Colibri and HXP-120 lamps provide more even illumination as conventional HBO mercury arc lamps, but this is still not sufficient for quantification. We measured 15% intensity difference between the best and worst illuminated areas. The fluorescence of the cells is measured by the intensity of their pixels. Objects with equivalent fluorescence will show different intensity depending on their position in the FOV and the illumination of that area. This error is corrected by a compensation image. Ten empty FOVs are recorded in random positions of the compensation slide in every channel. One final compensation image is created from the 10 images in every channel by omitting the darkest and brightest pixels and averaging the rest in every pixel position. This method eliminates local artifacts from the individual images which are usually darker or brighter than the compensation slide itself. The averaging lowers the noise of the camera (33). Every FOV image is compensated using the following equation:

  • equation image

I′ denotes the compensated image, I denotes the original image, and C denotes the compensation image. The x- and y-indices denote an image pixel, and u and v denote the coordinates of the brightest pixel of the compensation image. The method is described in detail with examples in our previous work (17).

To evaluate the effect of compensation 1,200 of the brightest Coulter beads were scanned with 20 ms exposure time and without compensation. Their CV value was compared with the same bead population scanned with compensation. These measurements were also used for the system linearity measurements described later.

Digital Gain

The AxioCam MRm grabs 12-bit images but the MIRAX system handles only 8-bit images. We implemented a Digital Gain function in the system which stores the user selected 8 bits from the original 12 bits. The default setting is digital gain = 0 which means that the most significant 8 bits will be stored and digital gain = 4 means that the least significant 8 bits will be stored. This way the user can select between scanning speed and image noise. Every step of the digital gain halves the exposure time but increases the noise. The recommended setting for standard imaging is digital gain = 2 because the exposure times are four times shorter and the increase in noise is hardly noticeable. For the quantitative measurements digital gain = 0 was used.

Image Segmentation

The system uses two level thresholding for image segmentation. An upper and a lower threshold can be defined and pixel values below and above those values will be excluded. Neighboring pixels are grouped to one object. Several segmentations can be included within one measurement and all of them are displayed by user selected colors on the digital slide. The image can be filtered before segmentation by Gauss, Median, or Wiener filters with a kernel size of 0– 10 pixels; 0 means that the filtering is off. The filter algorithms are implemented in the Intel Integrated Performance Primitives image processing package (Intel Corp., Santa Clara, CA). The actual image processing is done on the original pixel data, filtering is used only for segmentation. Detected objects can be further filtered based on their size in square millimeter (μm2). Segmentation settings can be saved as a Mirax Image Segmentation Profile file with .misp extension.

Quantitative and Stoichiometric Measurement

The following parameters are measured for each object: area, perimeter, shortest and longest diameter, shape factor, average pixel intensity, and integrated fluorescence (IF). Shape factor is calculated with the following equation:

  • equation image

The shape factor for a perfectly round object is 1. IF has the same function as fluorescence in FCM. The average value of the background around the object is subtracted from every pixel and the pixel values are summed (17). Area, perimeter, and diameters are measured in micrometer (μm). The system automatically calculates the μm/pixel value from camera type, objective magnification, and camera adapter magnification. This value is stored in the digital slide and can not be modified.

System Linearity Measurement

The system's linearity was verified by three different methods. To measure exposure linearity 1,000 beads from the brightest population were scanned with 5, 10, 15, 20, and 25 ms exposure time. To test single exposure linearity 1,800 beads were scanned at once from a mixture of all the five intensities. The IF of the different populations was compared with the manufacturer supplied intensity data. In both cases, the clustered beads were filtered from the measured objects.

To verify areal linearity, 5,300 beads and bead clusters from the third intensity were scanned and the IF of different cluster sizes was correlated.

Finally, we measured the CV value of Hoechst stained lymphocytes to assess usability on real samples.

Fluorescent Digital Slide

The standard proprietary file format of the MIRAX system was used to store the fluorescent digital slides. This format stores the images in an overlapped tiled format at 10 different magnifications. Every magnification layer is half size of the previous one starting from 1:1, 1:2, and down to 1:512. The storage of different magnifications is necessary to provide fast magnification change in the viewer software. One image of the camera is split into 4 × 4 smaller images because these can be handled more efficiently if the digital slide is browsed on the internet. Tiles can be stored internally in three different user selectable image file formats: BMP (Bitmap), PNG (Portable Network Graphics), or JPEG (Joint Picture Expert Group). The compression rate of the JPEG format can be defined by the user. Three fluorescent channels are stored in the red, green, and blue channels of a standard color brightfield image. If more than three channels are scanned, then a new layer of RGB image is stored in the file. The format also stores the preview image of the slide recorded by the preview camera and the label area image.

Fluorescent Virtual Microscopy

The MIRAX viewer software was used for digital slide display. The viewer has the following main functions: arbitrary magnification selection, panning, annotation handling, measuring and opening slides from a teleconsultation server on the internet. Additionally, several slides can be opened at the same time for comparison, annotation areas can be exported to reports. The viewer can be controlled from the keyboard for faster handling. Synchronized multiple participant teleconsultation sessions are also possible. The viewer stores which areas of the slide were examined at which magnification for quality control. Above the basic viewer functions, the fluorescent slides can be used with application packages which are optional software modules of the MIRAX viewer. The packages are implemented as standalone executable files but their functionality is closely integrated with the MIRAX viewer and they can be started only from the viewer. There are packages for tissue microarray analysis (TMA), quantitative measurement (HistoQuant), education (E-School), and three-dimensional (3D) reconstruction. The measurements in this study were made with the HistoQuant package.

The viewer had to be modified to read more than one image layer when more than three fluorescence channels are digitized. For every channel pseudo color, brightness, contrast and gamma can be set individually. These controls are on the bottom of the viewing area in Figure 4.

thumbnail image

Figure 4. MIRAX viewer. On the left side of the viewer window there are two previews. The upper one always shows the complete slide area for navigation. The lower preview shows a magnified image for better orientation. The main viewing area is in the center. In the upper right corner, there is a magnifier window showing the area around the mouse cursor with four times higher magnification as the main working area. Fluorescent channel selection, pseudo colorization, brightness, contrast and gamma settings are below the main viewing area.

Download figure to PowerPoint

Measurement Evaluation Tools

Measurements can be evaluated by the scatter plot, histogram, gallery and data export tools of the HistoQuant package. The scatter plot and histogram tools can display any measured parameters along their axis on linear or logarithmic scale. Data points can be gated and the gated data can be passed to other scatter plots, histograms, and galleries. The scatter plot tool supports square, ellipsis, and freehand gates. The package does not have statistical analysis functions. Gated data can be exported in a comma separated values file with .csv extension, which is interpreted by every spreadsheet or statistical program. We used Microsoft Excel for this purpose.

RESULTS

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

Focusing

The depth of field of the objective was calculated to be 0.57 μm at the 605 nm emission wave length of the Rhodamine filter (34). One step of the focus motor is 0.2 μm that is almost three times smaller as the depth of field thus the fine focus algorithm steps always in two motor steps (0.4 μm), because this is sufficient for obtaining well-focused images. Table 1 shows the differences between the autofocused and the manual focused focus levels.

Ninety-three percent of the FOVs did not differ at all or the difference was less than the depth of field. Some visual differences are observable within the depth of field but this does not influence the information content of the image. The observer might pick a different FOV as the autofocus algorithm.

In 7% of the cases, the difference was 0.8 μm. It means only one focus step error since the smallest step the algorithm makes is 0.4 μm. Subjectively assessing these images they are still well focused.

Scanning Results

The marker pen based sample detection worked reliably if the marking was at least 1-mm wide and continuous in the area imaged by the preview camera. The preview camera grabs 30 frames per second thus preview image capture time is negligible in the whole process.

Exposure times and focusing speed is proportional to the staining intensity. Table 2 summarizes the scanning information for five different samples.

Table 2. Scanning results for five different samples
Sample typeExposure time in milliseconds (ms) and digital gain setting in parenthesesArea (mm2)Number of field of viewsScanning time (h:min:s)Compression type and qualityFile size (MB)
DAPIFITCRhodamine
  1. The samples were scanned using the HXP-120 lamp and Zeiss Plan-Apochromat 20×, NA 0.8 objective with 0.32 μm resolution. Different sample types were selected to give an overview of the system's performance with regular fluorescent microscope slides and image cytometry samples.

TMA sample2 (2)35 (2)10 (2)5584,3861:43:57JPEG, 90%1,892
TMA sample1 (2)14 (2)2 (2)3042,39845:13JPEG, 90%814
Skin biopsy1 (1)18 (3)66 (0)1511210:31JPEG, 90%63
Coulter beads, third intensity400 (0)413256:42JPEG, 80%39
Hoechst control sample43 (0)252136:19JPEG, 100%87

Measurements: System Stoichiometric Linearity After Compensation

Figure 5 shows the graphs of the three linearity tests. In the first linearity test, the calculated correlation between exposure time and IF was 0.999491. The CV values at the different exposure times from 5 to 25 ms were 2.87, 2.72, 2.85, 2.54, and 2.43, respectively. The CV value of the uncompensated scan was 10.06.

thumbnail image

Figure 5. System linearity measured with Coulter beads. (A) The correlation between exposure time and integrated fluorescence (IF) of the brightest beads. The data points show IF. (B) The correlation between IF and the manufacture specified intensity of 1.5625%, 6.25%, 25%, and 100%. The data points show IF. (C) The correlation between the measured area of clusters and the number of beads in the cluster. Data points show area in square micrometer.

Download figure to PowerPoint

In the measurement of the bead mix, the unstained bead population was not visible. From the four stained population, the weakest is detectable but not reliably because the bead's pixel intensity values are around 3 and 4. The CV values from 1.5625% to 100% bead intensity are 10.53, 2.78, 1.89, and 2.05, respectively. The correlation between the manufacturer defined intensity ratios (1:4) and the measured IF ratios (1:3.60, 1:4.55, 1:3.57) was 0.999548. The scatter plot, histogram, and gallery of the brightest population are shown in Figure 6. We selected a subpopulation on the right side of the cloud that related to the perfectly round stand alone beads. We found that clustered beads do not appear as a group of perfect spheres, but on the point of contact they get distorted. The measurement of IF is based on the summation of the pixels and in the case of clusters, it does not provide precise results due to the deformation.

thumbnail image

Figure 6. Scatter plot, histogram, and gallery of a bead population in the HistoQuant program. Beads on a slide with five different intensities are scanned. (A) The scatter plot of the brightest beads. Shape factor is displayed versus area. Clustered bead are deformed and to measure intensity precisely the stand alone perfectly round objects are selected as a subpopulation on the right of the cloud. (B) The histogram derived from the scatter plot. The green marker lines delimit the final measured data set. (C) A portion of the bead gallery derived from the histogram.

Download figure to PowerPoint

Finally, we measured the areal linearity of the system. In this test, the brightest beads were used. The correlation between the average area of single, double, and triple clustered bead populations and the number of beads in a cluster was 0.999998.

In the Hoechst-stained control sample, the DNA fluorescence of 215 lymphocytes had a CV value of 6.4.

DISCUSSION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

Sample Detection and Localization

The MIRAX MIDI is an automated whole slide imager with a slide loading capability for 12 slides. To enable automated fluorescent scanning, it is important to develop an automated specimen detecting and locating method. Fluorescent samples have low contrast in normal brightfield imaging and they can not be detected by the preview camera. The possible alternative solutions could be low-magnification fluorescent, darkfield or phase contrast imaging or manual sample marking. The first three methods would require an objective changer to use a low-magnification objective to generate a slide preview. In fluorescent imaging, the intensity is calculated by the following equation:

  • equation image

Low-magnification objectives have low-numerical apertures that result in 1,000 times longer exposure times compared with the 20× objective used in this study, which is absolutely impractical. Darkfield illumination and phase contrast would require a different condenser design that could not be fitted in the MIRAX MIDI because of the tight construction of the quick focusing mechanism. Because of the optical and mechanical constraints, we decided to use the marker pen-based method which requires negligible extra burden from the users. The existing hardware could be used without any modifications, and it worked well on all samples.

Autofocusing

The required exposure time to image fluorescent stains by a microscope can vary from few milliseconds to several thousand milliseconds or in other words the system has to focus samples with a dynamic range of 1:1,000. The dynamic range of the AxioCam MRm is 12 bits but it was used in 8-bit mode because the complete MIRAX software package has 8-bit internal image handling.

The theoretical dynamic range of an 8-bit imaging device would be 1:255 but in practice the electronic noise and photon shot noise decrease it (33). The electrical noise of the AxioCam MRm was negligible in 8-bit mode. Photon shot noise equals to the square root of the photons arriving at a pixel. At low-light intensities, shot noise dominates the pixel differences and not the actual sharpness thus dark image are not sufficient for focusing. According to our tests if the brightest pixels were in the pixel value range of 32 and 64 then the calculated sharpness values were reliable for most samples. Low-contrast samples needed a range from 64 to 128 or from 128 to 255. We concluded that an 8-bit camera has a practical dynamic range of 1:8 (32:256) if images appropriate for focusing have to be grabbed. Our solution to find the right focus level of samples with a higher dynamic range was to continuously change the exposure time during the focusing process. A fixed long exposure time would saturate the sample and a short exposure could miss faint samples.

The brightfield sharpness calculation algorithm of the MIRAX software worked well on fluorescent tissue samples. However, on cell samples where only few cells were in the FOV the algorithm did not work reliably. The brightfield focus algorithm calculates and sums the value differences of neighboring pixels. A lymphocyte has about 7 μm diameter that is represented roughly by 370 pixels with 0.3225 μm/pixel resolution. 370 pixel is 0.026% of the active pixels of the camera and the pixel differences of the cell disappear in the sum of differences of the background noise.

Our goal was to create a general algorithm without a fixed threshold to separate the background from the samples. By calculating the fifth power of every difference we could enhance the higher differences with lower occurrence and suppress small differences with high occurrence. This exponential value was determined by experiments and worked reliably on all samples. Figure 7 shows a FOV in two different focus positions and the calculated sharpness values.

thumbnail image

Figure 7. Sharpness calculation. (A, B) A fluorescent field of view in and out of focus by 4 μm. Surprisingly, the summed pixel value differences of both images are 126,588. The sharpness value which sums the fifth power of every pixel value difference is 3.01 × 1010 for the sharper and 3.28 × 108 for the out of focus image. From a zero difference, the sharpness algorithm created a difference of two magnitudes which was possible because the sharper image originally had more higher pixel value differences.

Download figure to PowerPoint

Quantitative Imaging

Illumination

The HXP-120 fiber coupled metal-halide short arc lamp has several advantages over traditional HBO mercury arc lamps. It has more light, five times longer life time and due to the fiber coupling the illumination unevenness is not noticeable visually on many samples. Unfortunately arc lamps suffer from arc wandering. The arc between the electrodes of the lamp jumps from time to time to another point on the electrodes and change the illumination pattern and intensity (http://zeiss-campus.magnet.fsu.edu/tutorials/arclampinstability/index.html). If the illumination pattern changes the prerecorded compensation images are not sufficient anymore. We found that the mechanical position of the fiber can also change the illumination pattern. These properties make the metal-halide lamp impractical for quantitative measurements. The Colibri lamp provided very stable illumination in time and it is mounted directly on the microscope without a fiber. These features made this lamp the optimal choice for quantitative analysis.

The power of the two light sources is not directly comparable because their spectral characteristics are different but our experience was that the Colibri was 20–30% weaker on average. The meta-halide lamp emits light from 300 to 700 nm with the same peak wavelengths as HBO lamps (365, 405, 436, 546, and 579 nm). The peaks are weaker compared with HBO but between the peaks there is more light. LED modules for the Colibri lamp are available with 365, 380, 400, 455, 470, 505, 530, 590, 615, and 625 nm center wavelengths.

Compensation slide

Previously, we used slides with FITC solution to record compensation images (17). The manually prepared slides contained bubbles, dust particles, and other artifacts therefore proper FOVs had to be selected manually. The thin polymer section used in this study covered the complete slide and we did not detected irregularities. The described compensation image creation algorithm always provided good results. The laser dyes in the compensation slide fade. Our algorithm picks only 10 FOVs randomly thus slides can be used several times. We experienced that it is important to focus the compensation slide to map the illumination pattern as good as possible. The FITC solution was easy to focus because of the mentioned artifacts. On the polymer slide, the focus could not be determined based on the sample structure as there was none. The slide was focused by selecting the motor position which provided the brightest image based on the properties of exposure and focus shown in Figure 2.

Image segmentation and measurement

HistoQuant is designed as a general purpose image processing application and as such manually set threshold values are necessary for segmentation because biological samples are complex and an automated threshold algorithm cannot determine the goal of a study. For example nucleus, cytoplasm, different cell populations, and FISH spots require different threshold values. The possibility to set different threshold values within one measurement for the different targets is also necessary. In the linearity measurement of beads with different intensities, we used the following lower and upper threshold values for the four populations: 4/7, 6/17, 15/70, and 60/255. If all four populations would have a threshold with a single pixel intensity level of 4, then the bright beads would have ∼2.7 times larger area because the background around a bright object is also brighter. Upper thresholds are also practical to better discriminate the different groups. One threshold value was sufficient for areal linearity measurements, exposure linearity measurements and the lymphocyte measurement for the samples with homogenous fluorescence intensity.

Manually selectable threshold values would not be optimal for use in a high throughput system in a clinical routine environment. In the literature, there are several publications on automatic detection and classification of nuclei and FISH analysis on fluorescent images (35–37) and complex automated segmentation methods for membrane and muscle fiber for example (31, 38).

In our measurements, CV values of the beads varied between 1.89 and 10.53. The average CV value in the useful intensity range without the outlier 10.53 value was 2.52. The bead manufacturer measured CV values between 0.7 and 0.3 at PMT voltages from 400 to 650 V, respectively, on a Beckman Coulter FC 500 flow cytometer (39). At 300 V, the CV increased to 4.2. The average CV value in the useful range was 0.5. Goller and Kubbies (40) used two different UV lasers to measure with FCM the CV values of peripheral blood lymphocytes stained with the same Hoechst 33258 stain that we used. Their results varied between 3.9 and 4.3. Stokke and Steen (41) measured Hoechst 33258 stained human blood leucocytes with an arc lamp based flow cytometer. They measured CV values ranging from 3.0 to 3.6 depending on the different fixation and treatment techniques. Comparing these results with our measured CV value of 6.4 on mononuclear blood cells, we can conclude that our imaging system measures higher CV values by an additional 2–2.7 on average. The increase is probably caused by the compensation because the illumination pattern changes from slide to slide due to differences in mounting medium, sample thickness, cover slip, and slide material. These higher values are still acceptable for cytometric work.

CONCLUSION

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

In our previous work (17), we showed that it is possible to use a regular fluorescent microscope as an image cytometer. The disadvantage of that solution was the complex and inconvenient work flow and the possible instability due to the HBO lamp and manually prepared compensation slides. The use of new automated whole slide imager mechanics, stable LED light source, and high quality compensation slides can solve these problems. WSI has many advantages in archiving, telepathology, image processing, education, and measurements for routine and research use. Since WSI can be used virtually in every field of light microscopy, there are many more available resources in the industry for the development of the technology. With our developments, the cytometric community can also benefit from these advancements but to apply this technology in the clinical routine the development of automated sample specific image processing applications is necessary.

Acknowledgements

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED

The authors thank Ferenc Szipőcs for correcting the text.

LITERATURE CITED

  1. Top of page
  2. Abstract
  3. INTRODUCTION
  4. MATERIALS AND METHODS
  5. RESULTS
  6. DISCUSSION
  7. CONCLUSION
  8. Acknowledgements
  9. LITERATURE CITED