Abstract
Simultaneous detection of two different antigens on paraffin-embedded and frozen tissues can be accomplished by double immunohistochemistry. However, many double chromogen systems suffer from signal overlap, precluding definite signal quantification. To separate and quantitatively analyze the different chromogens, we imported images into a Macintosh computer using a CCD camera attached to a diagnostic microscope and used Photoshop software for the recognition, selection, and separation of colors. We show here that Photoshop-based image analysis allows complete separation of chromogens not only on the basis of their RGB spectral characteristics, but also on the basis of information concerning saturation, hue, and luminosity intrinsic to the digitized images. We demonstrate that Photoshop-based image analysis provides superior results compared to color separation using bandpass filters. Quantification of the individual chromogens is then provided by Photoshop using the Histogram command, which supplies information on the luminosity (corresponding to gray levels of black-and-white images) and on the number of pixels as a measure of spatial distribution.
Double immunohistochemistry has been developed to allow the simultaneous detection of different antigens on cells contained on a tissue section. This technique is particularly important in situations where the immediate proximity of different cells or co-expression of different antigens on certain cells is being investigated. In addition, double immunohistochemistry is used when information on the spatial contribution of tissue elements (e.g., tumor cells vs stroma) is sought. Over the past few years the techniques of double immunohistochemistry have been refined considerably (van der Loos et al. 1993, 1996). Differential antigen recognition is made possible using different enzymatic tracers, such as β-galactosidase, which gives a turquoise color, and alkaline phosphatase, which gives a bright red color. These colors can be distinguished easily for qualitative estimation of the distribution of cell populations on a given slide. To quantify the information contained in the cell- and tissue-specific chromogen distribution on the microscope slide, a differential analysis of the color distribution becomes necessary. For that purpose, colored filters have been applied, selectively excluding the transmission of light of a certain wavelength (Zhou et al. 1992). However, these techniques can be applied only to chromogens with a clearly distinct, nonoverlapping light spectrum, and even then cannot discriminate completely enough among colors, which always exhibit a certain range of hues and luminosities within a tissue section.
Photoshop, a program developed and constantly refined for the manipulation of digitized images, has become one of the essential tools in graphic design, desktop publishing, and advertising agencies. To allow manipulation of selected areas or features contained within the digitized image, Photoshop supplies sophisticated tools and commands for the recognition, selection, and separation of objects, shapes, gray levels, and colors. We show in this report that this latter feature of Photoshop can be applied to the differential selection and quantitative analysis of chromogens on digitized images taken from double immunostained microscope slides.
Materials and Methods
Antibodies and Reagents
Mouse anti-CD68 (clone PG-M1), mouse smooth muscle anti-α-actin (clone 1A4), mouse anti-vimentin (clone V9), normal goat serum, alkaline phosphatase-conjugated streptavidin (strep/AP), biotinylated goat anti-mouse immunoglobulin (GAM/biotin), normal mouse serum, rabbit anti-fluorescein isothiocyanate (FITC), alkaline phosphatase-conjugated goat anti-rabbit immunoglobulin (GAR/AP), and New Fuchsin AP substrate system were from DAKO (Glostrup, Denmark). FITC-conjugated mouse anti-cytokeratin (clone CAM5.2) was from Becton Dickinson (San Jose, CA). Alkaline phosphatase-conjugated goat anti-mouse IgG3-specific (GAM-IgG3/AP), biotinylated goat anti-mouse IgG2a-specific (GAM-IgG2a/biotin), and β-galactosidase-conjugated goat anti-rabbit immunoglobulin (GAR/GAL) were from Southern Biotechnology (Birmingham, AL). β-Galactosidase-conjugated streptavidin (strep/GAL) and 5-bromo-3-chloro-2-indolyl-β-D-galactopyaranoside (X-GAL) were from Boehringer Mannheim (Mannheim, Germany). Naphthol-ASMX phosphate, Fast Blue BB, ferrihexacyanide, and ferrohexacyanide were from Sigma (St Louis, MO). Antisera and antibody/enzyme conjugates were diluted in Tris-HCl (50 mM, pH 7.8)-buffered saline (TBS) + 1% bovine serum albumin (BSA). TBS washings were performed between all steps (three times for 2 min) and all incubations were performed at room temperature (RT) unless otherwise stated.
Immunohistochemical Technique
Human carotid artery segments with atherosclerotic lesions were fixed in phosphate-buffered formaldehyde for 24-72 hr and then embedded in paraffin. Five-μm-thick sections were mounted on organosilane-coated slides and dried overnight at 37C. Sections were deparaffinized in xylene, rehydrated in graded alcohols, and washed with tapwater. Then the sections were treated for antigen retrieval using citrate (10 mM, pH 6.0) in a household microwave oven (Cattoretti et al. 1993): 3 min at 900 W to bring to boiling, 15 min at 250 W to maintain boiling, and 10-min cool-down in the citrate buffer outside the oven. All sections were washed with TBS (three times for 2 min) and covered with normal goat serum (1:10, 15 min).
Breast carcinoma specimens were snap frozen in isopentane-cooled liquid nitrogen and stored at −80C. Five-μm sections were cut, dried overnight under a fan at room temperature, and either used right away or stored dry in a closed box at −80C. Cryostat sections were acetone-fixed (10 min, 4C), briefly air-dried, and extra fixed to preserve a better tissue morphology with Zamboni's fluid (picric acid/paraformaldehyde in phosphate buffer, pH 7.4) for exactly 2 min, washed with TBS (three times for 2 min), and then covered with normal goat serum (1:10, 15 min). Double labeling was based on either two primary antibodies from different IgG subclasses (CD68/α-actin; Figure 1), or one unlabeled primary antibody and one FITC-conjugated antibody (vimentin/cytokeratin; Figure 2) (van der Loos et al. 1993, 1996).
From the available color combinations for double staining, we selected one with a superior color contrast: turquoise/red, as obtained with β-galactosidase and alkaline phosphatase as tracer enzymes, and X-Gal and New Fuchsin as chromogens, respectively (van der Loos et al. 1987). This color combination can be applied for recognizing two different cell types (van der Wal et al. 1994) and for revealing co-localization by a purple-blue mixed color showing contrast to the basic colors (Naruko et al. 1996; van der Loos et al. 1996). The following subsequent incubations were performed for the red/turquoise CD68/α-actin double staining applied on paraffin sections from the carotid artery (Figure 1): mouse anti-CD68, PG-M1 (IgG3) (1:50) in a cocktail with mouse anti-α-actin, 1A4 (IgG2a) (1:100, 60 min); GAMIgG3/AP (1:20) in a cocktail with GAM-IgG2a/biotin (1:100, 30 min); strep/GAL (1:40, 30 min). The following subsequent incubations were performed for the red/turquoise vimentin/cytokeratin double staining applied on paraffin sections from breast carcinoma (Figure 2): anti-vimentin (1:200, 60 min); GAM/biotin (1:200, 30 min); cocktail of strep/AP (1:100) and normal mouse serum (1:10, 30 min); FITC-conjugated anti-cytokeratin (1:20, 60 min); rabbit anti-FITC (1:1000, 15 min); and finally GAR/GAL (1:10, 60 min).
Double staining specific controls consisted of replacing either one of both primary antibodies by nonimmune reagents of the same species or IgG subclass, keeping the protein concentration similar to the primary antibodies. After all antibody detection steps the enzymatic activities were visualized. For the red/turquoise color combination, first β-galactosidase activity was developed with X-GAL and ferro-ferricyanide (Bondi et al. 1982) and next, after washing with TBS, AP activity was developed in red with the DAKO New Fuchsin AP substrate system according to the instructions of the manufacturer. Cryostat sections were postfixed with buffered formaldehyde 4% after the enzymatic visualization. Sections were finally rinsed with distilled water and aqueously mounted with Ultramount (DAKO).
Image Analysis
Images are imported into the S-VHS port of a personal computer (G3 Power Macintosh with inbuilt graphic capture board) using a one-chip CCD red-green-blue (RGB) color video camera (JVC TK-C1381 camera; Tokyo, Japan) and a standard diagnostic microscope (Dialux 22; Leitz, Wetzlar, Germany) equipped with a halogen light source connected to a stabilized, adjustable power supply (12 V, 100 W). Images are opened in Photoshop (version 4; Adobe Systems, San Jose, CA) and stored in a Photoshop or a PICT file format on the hard drive or on an external data storage device (ZIP drive; Iomega, Roy, Utah).
Chromogen separation by bandpass filters (
Cryostat section from an invasive breast duct carcinoma. The carcinoma section is composed of epithelial tumor cells (cytokeratin, alkaline phosphatase, New Fuchsin) and stroma (vimentin, β-galactosidase, X-Gal). In
The technique of selection of similar features on a digitized immunohistochemical image has previously been described in detail (Lehr et al. 1997). The selection of a specific color on a digitized image is a “procedural” selection, i.e., using the information of color saturation, hue, and luminosity intrinsic to every pixel in the image. To make a selection of a specific color, the cursor of the Magic Wand tool is clicked on any object on the image displaying the desired color/chromogen. The selected area is automatically highlighted. To specify how broad a range of color the Magic Wand tool should include in the selection, the Tolerance value in the Magic Wand Options palette can be set to a number between 0 and 255, with lower numbers indicating a small range of colors. Using the Select Similar command, all pixels on the image are highlighted that fall within the selected color range and are not touching the initial selection. Because the selected area is automatically highlighted on the screen, the selection process can be controlled at every step and necessary corrective measures can be taken. For example, the Select Grow command expands the color range to expand the selection to neighboring pixels. The selection continues to grow as often as this command is repeated. Often colors/chromogens are too similar to allow a control of separation using the above-described method. In this case, tone adjustments must be made to the original image. Using the Hue/Saturation tool in the Image Adjust menu, the saturation and/or hue of a selected color (red, yellow, green, cyan, blue, magenta, or any color selected using the Eye-dropper tools) can be modified (e.g., rendering red into a bright pink color, which now separates well from the turquoise chromogen; see Figures 1B, 1E, 2A, 2B).
Once the different chromogens are selected, quantification is accomplished using the Histogram command in the Image menu. This display is rarely if ever used by graphic designers but rather serves as an internal measurement of tonal distribution as the basis for automated image manipulation (map commands). When Histogram is selected, a display appears on the screen depicting the gray levels (black/white) or the luminosity (color) of all pixels within the selected area, including median and standard deviation. Furthermore, this display shows the number of pixels that are covered by the selected area. Because the number of pixels reflects a surface area on the image, important spatial information can be obtained for the specific chromogen (and hence the cells expressing a certain antigen) and can be expressed as percentage of the entire image or in μm2.
Results
We demonstrate in this study that complete chromogen separation is possible using Photoshop-based image analysis. In contrast to color separation by the use of bandpass filters, which only incompletely eliminated the respective chromogen on the image (Figures 1A-1D), Photoshop-based image analysis was found to completely separate the two chromogens (Figures 1E-1G). When applied to breast cancer tissue composed of epithelial cells (cytokeratin antibody, New Fuchsin, red chromogen) and stroma (vimentin antibody, β-Gal, turquoise chromogen), we found that color separation was complete. The surface areas of epithelial cells and stroma, assessed as numbers of pixels covered by the respective chromogens using the Histogram tool, added up to almost 100% of the entire tumor field (70.2 + 30.3%; Figure 2). A similar result was obtained when chromogen separation and quantification were performed on a random high-magnification section of the tumor (Figure 2, insets), demonstrating that Photoshop-based image analysis allowed complete chromogen separation irrespective of the image magnification.
Discussion
The present study demonstrates that complete chromogen separation is possible on digitized multispectral images using Photoshop-based image analysis. This software program makes use of the integral information on color hue, saturation, and luminosity of every individual pixel in a digitized image. To better visualize and hence to control the efficacy of color separation, manipulations of saturation and/or hue can be performed on the different colors contained in the image. This is best demonstrated in Figures 1B and 1E and in Figures 2A and 2B, in which the luminosity of the turquoise chromogen is enhanced and the red chromogen is turned into a bright pink color. Although this manipulation does not affect the ability of the software to separate the colors, it does help to control the effective separation of the two chromogens on the computer monitor.
Over the past decade, color separation on digitized images of immunohistochemical slides has been performed successfully using diverse custom-made image analysis programs such as IBAS 2000 (Deverell et al. 1989), BQ MEG IV Vista color system (Shapiro et al. 1992), MAPPS-II (Goto et al. 1992), VIDAS (Willemse et al. 1993), Quantimed 600 (Kohlberger et al. 1996), COSAS (Black and Rosen 1996), MAGICSAN (Mosedale et al. 1996), and NIH image (public domain program; Ruifrok 1997). In these previous studies, color separation was performed by automatically thresholding red-, green-, and blue-filtered gray-scale values of the image. This technique was applied to separate and differentially analyze diaminobenzidine (brown)-stained antigen-positive cells/areas from hematoxylin (blue)-counterstained cells/areas (Shapiro et al. 1992; Kuyatt et al. 1993; Ruifrok 1997) and to differentiate and quantify triphenyltetrazolium chloride (TTC)-negative infarcts from noninfarcted brain tissue in experimental cerebral infarction in rats (Goldlust et al. 1996). In essence, this technique works in analogy to bandpass filters placed over the images. For this reason, we compared the Photoshop-based image analysis to a technique in which two narrow bandpass filters were layered over the color image. Depending on the spectrum of the red (alkaline phosphatase, New Fuchsin) and blue chromogen (β-galactosidase, X-Gal), the filters were chosen so as to reduce the spectral overlap as much as possible (Figures 1A-1D). A similar technique has been described previously by Zhou and coworkers (1992) for the differential analysis of diaminobenzidine (brown)-stainedantigen-positive cells from hematoxylin (blue)-stained background, although in that study the images were digitized in black and white and the relative gray levels of the antigen-positive cells and the background were quantified. However, Figure 1 also demonstrates that the bandpass filters only incompletely eliminated the respective chromogens on the image (Figures 1C and 1D), resulting in a marked methodological inaccuracy compared to the Photoshop-based color separation in Figures 1F and 1G.
In further developments of the described techniques, RBG thresholding was improved considerably by integrating into the thresholding algorithm information on hue, luminosity, and saturation (Gato et al. 1992; Lamaziere et al. 1993; Smolle 1996). When the spectral characteristics of the image components do not change over the area of a slide or from slide to slide, constant weighted linear combinations of spectral images can be used to generate one-dimensional or two-dimensional images that provide a contrast between the image components. However, in histological or immunohistochemical slides, the spectral characteristics are rarely constant, i.e., they vary with the histological technique and, above all, with the extent of epitope availability for reaction with a respective antibody/enzyme/chromogen complex (MacAulay et al. 1989). Although various chromogens used as optical stains are perceived as specific colors, they in fact have complex attenuation spectra. As a consequence, attenuation from several colors contributes to the overall hue, saturation, and luminosity, and color separation can no longer be reliably achieved using a fixed, predefined threshold technique (Zhou et al. 1992). This problem is easily overcome on the Photoshop-based image analysis, because the desired chromogen/chromogen range is selected for each individual case (by simply clicking a representative area with the Magic Wand tool) and separated using the Select Similar mode based on the combined information on color hue, saturation, and luminosity integral to the digitized image.
As a practical example, we applied Photoshop-based color separation to the selective quantification of breast carcinoma tissue (Figure 2) covered by epithelial tumor cells (cytokeratin antibody, New Fuchsin, red chromogen) or by stroma (vimentin antibody, β-Gal, turquoise chromogen). We found that, in addition to complete chromogen separation, the calculated surface area of epithelium and stroma added up to almost 100% of the selected tumor field. Of particular interest is the fact that color separation is effectively performed at both low (Figure 2) and high magnifications (Figure 2, insets). In contrast to RGB thresholding techniques, which demonstrate improved separation capacities at higher magnifications (Brown et al. 1998), Photoshop-based image analysis uses the color information contained in each individual pixel (color hue, saturation, luminosity) and thus allows separation of colors at the levels of pixel resolution. This also makes it possible to completely separate chromogens with closely related spectral characteristics, which is particularly relevant to the separation of brown and blue for the separation of the immunoperoxidase signal from the hematoxylin counterstain. We are now using this technique in a study designed to compare the mitotic activity in breast cancer with the proliferative activity assessed immunohistochemically using the Ki-67 antigen. In this study, the number of MIB-1-positive tumor cells is counted using Photoshop-based image analysis as previously described (Lehr et al. 1996) and is expressed per area of tumor epithelium. This technique will help to better assess the proliferative activity of breast cancers with largely different proportions of tumor epithelium and stroma (or mucin lakes, in the case of mucinous carcinomas).
There are only a few reports in the literature in which Photoshop has been applied to biomedical research. Several authors have used Photoshop merely as a way of importing scanned images into their computers for later analysis by other programs (Brown et al. 1995; Westhuyzen et al. 1997). For example, Ikeda and coworkers (1997) have used Photoshop not only to scan selected fields from microscope slides but also to compose seamless widefield images of large tissue sections. Others have used Photoshop to superimpose grids over scanned images (Gatlin et al. 1993) or to assess facial movement as a means of grading facial nerve injury (Sargent et al. 1998). However, none of these studies has actually made use of the tools for recognition and quantification described here. We have previously applied Photoshop-based image analysis to the quantification of proliferative activity (using MIB-1) and of hormone receptor expression in invasive breast cancer (Lehr et al. 1996, 1997), and now to the color separation in multispectral images (this study). In addition to the ease of use and the versatility of Photoshop-based image analysis, one key advantage of the described technique is the widespread accessibility of the program, which can be acquired for a fraction of the costs of custom-made professional programs or is even supplied for free as accessory software for image acquisition with the purchase of many flatbed or slide scanners. We should also stress that some of the features used for color separation and for the image analysis described previously (Lehr et al. 1996, 1997) and in this report are also integral components of many other imaging programs.
Footnotes
Acknowledgments
We are indebted to Drs Frans Nauwelaers and Peter Oud for the idea of applying a narrow bandpass filter to the digitized images.