"Visualization of Microscopy-Based Spectral Imaging Data from Multi ...

Visualization of Microscopy-Based Spectral Imaging Data from Multi-Label Tissue Sections

UNIT 14.19

James R. Mansfield,1 Clifford Hoyt,1 and Richard M. Levenson1 1

Cambridge Research & Instrumentation (CRi), Woburn, Massachusetts

ABSTRACT Combining images taken with light of specific wavelengths can dramatically enhance lightmicroscopic images. This technology is enabled by the availability of programmable filters that can be set to transmit light only of particular wavelengths. Spectral imaging technologies have become an important part of microscopy, and are particularly useful for analyzing samples that have been labeled with multiple (two or more) molecular markers. The most commonly used methodology for separating the markers from each other is linear unmixing, which results in a quantitative image of the location and amount of each marker present in the sample. The very complexity of these multilabel samples requires a high degree of sophistication in methods to visualize the results of unmixing. This article describes a wide range of useful visualization tools designed to better enable discrimination of different features in multilabeled tissue or cell samples. These commercially available tools can be attached to the standard laboratory light microscope to significantly enhance the power of light microscopy. Curr. Protoc. Mol. Biol. C 2008 by John Wiley & Sons, Inc. 84:14.19.1-14.19.15.  Keywords: spectral imaging r unmixing r data visualization

INTRODUCTION The application of gene and protein array methods for analysis of homogenized samples, and of flow cytometry for phenotyping isolated cells, has led to a much better understanding of the roles and interactions of multiple molecular species in normal and diseased tissues. However, in many cases, it is important not just to measure overall expression levels of specific molecules, but also to capture their spatial distribution in a relatively intact cellular and tissue architectural context. Multianalyte immunohistochemistry, whether in bright-field or fluorescence, has many potential applications in cell biology, clinical pathology, and, particularly, in preclinical drug development, in which multiple pathways implicated in drug action or drug resistance have to be interrogated. Typically, the proteins and other molecules of interest will be present in the same or spatially overlapping cellular compartments. Such co-localization (which does not necessarily imply intimate biochemical interaction) complicates imaging-based methods for multiplexed detection and quantitation, especially when spectrally overlapping labels are used.

However, a technique that combines imaging and spectroscopy (“spectral imaging”) can largely overcome these challenges by resolving overlapping chromogenic or fluorescent labels, thereby generating quantitative images of individual analytes. Also, through its ability to disentangle signals of exogenous fluorescent labels from the sometimes dominant endogenous autofluorescent background, in many studies spectral imaging can greatly improve sensitivity and quantitative accuracy. Spectral imaging is an ideal complement to multiplexed labeling strategies made practical by automated staining platforms; together these tools constitute clinically viable techniques that, in particular, hold great potential for prognostically and therapeutically profiling individual cancers.

INSTRUMENTATION An imaging spectrometer acquires the spectrum of each pixel in a two-dimensional scene. The conceptually easiest way to accomplish this is to acquire separate images of the scene, each at a different wavelength, and then to “stack” this image data in computer memory so that each spatial location in the scene is

Current Protocols in Molecular Biology 14.19.1-14.19.15, October 2008 Published online October 2008 in Wiley Interscience (www.interscience.wiley.com). DOI: 10.1002/0471142727.mb1419s84 C 2008 John Wiley & Sons, Inc. Copyright 

In Situ Hybridization and Immunohistochemistry

14.19.1 Supplement 84

Visualizing Microscopy-Based Spectral Imaging Data

associated with its brightness value at every acquired wavelength. Optomechanically, this is readily accomplished simply by rotating a filter wheel in front of a grayscale digital camera and then snapping a picture as each filter is positioned in front of the lens. While simple in concept, this approach becomes untenable as the number of wavelengths to be acquired grows above a relatively small number. In addition, there are drawbacks involved in actually rotating the filter wheel, including noise, vibration, and the potential for image shift if the filters are not all perfectly flat with respect to the plane of the camera. Fortunately, the same result can be achieved using electronically tunable filters. As the name implies, these devices can tune their spectral “pass-band” electronically and without moving parts. Advantages include quiet and vibration-free operation, high switching speed, and spectral flexibility, since one is not limited to the filter choices installed in a filter wheel. There are a variety of tunable filter designs; the examples in this article were all acquired using an imaging system based on the use of liquid crystal tunable filters (LCTFs). These are small, low-power, highly stable devices that can be set to transmit light across a range of wavelengths in the visible and extending into the near-infrared. The Nuance spectral imaging system (a registered trademark of Cambridge Research and Instrumentation, CRi, http://www.cri-inc.com) combines an LCTF with a low-noise megapixel CCD camera, optics to ensure parfocality (i.e., images in focus through the eyepieces are also in focus on the camera), and appropriate software to permit easy multispectral image acquisition and analysis. Since this device is equipped with a female C-mount connection, it can be installed on any microscope capable of supporting a standard digital camera. More complex and expensive ways of achieving spectral imaging capabilities have become popular with the widespread adoption of multispectrally enabled laser-scanning confocal microscopes, now available from all major microscope companies (Nikon, Olympus, Zeiss, and Leica) and frequently found in many core imaging facilities. This growth of spectral imaging (also termed “multispectral” or “hyperspectral” imaging, depending on the number of wavelengths used) has lead to a similar growth in the number and type of applications for which spectral imaging is used. These applications range from live cell imaging (Hiraoka et al., 2002; Zimmerman et al., 2003) to bright-field immunohistochemistry

(Levenson and Mansfield, 2006) to imaging in whole animals (Levenson et al., 2008). Multispectral laser-scanning confocal (or two-photon) systems are best suited for applications in which the confocality they provide is important, either for optically resolving subcellular features along the z axis or for attempting to view signals beneath the surface of thick samples. However, if optical sectioning is not absolutely required (which is the case for the majority of microscope-based imaging), then multispectral discrimination may most effectively be achieved using non-confocal approaches because of their much lower cost, lack of dependence on lasers, lower complexity, and superior light budget. However, multispectral imaging in general is difficult to deploy for some live-cell applications requiring high-speed acquisitions, since the technique is typically more time-consuming than singlewavelength-based methods. Spectral resolution requirements for good spectral separations, somewhat surprisingly, may not be that demanding; in general, systems that have a spectral resolution that is approximately half the bandwidth of the fluorophores or chromogens being used will maximize signal-tonoise without sacrificing separation ability (Neher and Neher, 2004). A full review of these methods is beyond the scope of this paper and can be found elsewhere (Garini et al., 2006).

LABELING One critical aspect of obtaining data from multiple molecular markers in a single tissue section is the method used for introducing molecular labels into the sample. Indirect antibody labeling is the most commonly used technique, but the complexities of managing multiple primary and secondary antibody combinations can be overwhelming. In fluorescence, direct labeling using antibodies conjugated directly to fluorescent labels is conceptually straightforward (Robertson et al., 2008). Using this methodology, multiple fluorophores can be introduced into the sample, either simultaneously or sequentially, and imaging can be performed after a simple washing step. However, this approach is less sensitive than indirect labeling methods (because of the latter’s opportunity for signal amplification), and much less flexible. Also, very few of the thousands of available antibodies can be purchased already labeled, so investigators must do the conjugation themselves, adding complexity and considerable cost.

14.19.2 Supplement 84

Current Protocols in Molecular Biology

In the case of bright-field chromogenic labeling, immunoenzyme methods are required, and for dually (Tacha and Miller, 2004; Hameed and Humphrey, 2005), triply (van der Loos et al., 1987; Herawi and Epstein, 2007), or even quadruply labeled samples, care must be taken to ensure that there is no crossreactivity between antibodies or that one chromogenic deposition does not block further depositions in the same spot (van der Loos et al., 1993). More generally, it is essential to pay attention to fixation and antigen-retrieval conditions, since certain analytes that one may wish to image simultaneously may require incompatible tissue preparation methods. An advantage that spectral imaging provides is that multiple chromogenic labels that may have been too similar in color to be easily distinguished by eye can now be used together (van der Loos, 2008). In general, it is highly recommended that a series of controls of all of the combinations of markers be used to check for labeling accuracy, antibody cross-talk, and interference between one label and another. Although this can be a great deal of work for a sample with many markers, it is an important step in methods development, and can make the determination of spectral libraries (for unmixing) easier and more accurate. Success in separating similar spectral species depends on the number of species to be distinguished, their absolute and relative brightness, the dynamic range of the camera, and the presence or absence of noise or other confounding spectral contributors (such as autofluorescence), as well as whether the species are physically separated or are actually co-localized in the same pixels. That said, by way of example, dyes that differ by just 2 to 3 nm in peak wavelength can be readily distinguished under favorable, but achievable, microscopy conditions using a standard Nuance imaging system. Again, the total number of signals that can be separated depend on all of the variables listed above, plus whatever difficulties may be encountered in actually achieving good-quality, highly multiplexed labeling (Neher and Neher, 2004).

SPECTRAL IMAGING METHODS Multicolor fluorescence and bright-field samples containing up to six or eight molecular markers can be successfully imaged and unmixed into separate channels. In general, most implementations employ a linear unmixing strategy for separating the contributions of

the various markers (Dickinson et al., 2001; Zimmerman, 2005). Once the signals are separated, there are different approaches for visualizing the results, enabling the spatial and quantitative relationships of the biological signals in the samples to be appreciated. This paper will cover a variety of methods for extracting this biological information, starting from the means by which the marker levels are calculated, and then focusing on visualization methods. This is not intended to be an exhaustive listing of methodologies, and other visualization methods exist (Rothmann et al., 1998; Rueden et al., 2004; Jaskolski et al., 2005).

Linear Unmixing Methodology One of the major data processing issues for spectral imaging is the separation of a mixture of spectral signals from each other so they may be studied independently. In principle, this separation of signals is termed a deconvolution, which is exactly what it sounds like: the undoing of a convolution, or mixing (Press et al., 1992). The majority of deconvolution methods in biological imaging are directed towards sharpening blurry images or increasing depth, or z-axis resolution, of three-dimensional image sets (Swedlow, 2007), and there are a great many mathematical approaches for this. However, for spectral imaging data, the process of separating the contributions of the individual fluorophores or chromogen signals from one other is usually performed with some form of a least-squares fit, or regression (Kariya and Kurata, 2004). In this technique, experimental data obtained from observations are approximated by adjusting the parameters of a model to give an optimal fit—i.e., adjusting the coefficients of a linear combination of the amount of each fluorophore or chromogen spectrum that, when added together, will best reconstitute the measured spectrum. This methodology is typically referred to as linear unmixing or simply “unmixing” (Zimmerman, 2005). In matrix notation, the normal equation for a least-squares fit of a single variable is written as:

(X X ) β = X T

T

y

Here, X is the data matrix of spectra, XT is its transpose, and y is the dependent variable. This is a simultaneous set of linear equations ˆ the least-squares that can then be solved for β,

In Situ Hybridization and Immunohistochemistry

14.19.3 Current Protocols in Molecular Biology

Supplement 84

estimators of the parameter values:

(

β = X TX

)

−1

XTy

For most actual data, however, it is not good practice to invert the normal equations matrix, due to numeric instabilities. Commonly, these equations are solved using the QR decomposition method (Horn and Johnson, 1985) or singular value decomposition method (Lawson and Hanson, 1974). For spectral imaging data, an optical spectrum is available at each pixel of the image and, usually, no measured dependent variable for each pixel. In this case, it is often convenient to think of a linear model where, by adjusting the fit coefficients, we minimize the residual between the data spectrum, S, and the fitted model:

(

)

min || Sij − aij A + bij B + ... ||2

Here: Sij is the measured sample spectrum at pixel i,j; A, B, and so on are the basis set, or spectral library, being used for unmixing; aij , bij , and so on are the least-squares fit coefficients at pixel i,j.

Visualizing Microscopy-Based Spectral Imaging Data

The values of aij and bij returned by the leastsquares fit represent images which are the “deconvolved” results of separating the contribution of each of the components from the others. These “unmixed” images form the basis for all of the data analysis of the spectral imaging dataset. Most, if not all companies that sell spectral imaging systems, regardless of the spectral discrimination technology used, incorporate some form of linear unmixing into their software. Many manufacturers have included refinements to their linear unmixing algorithms, such as non-negativity constraints on the fit coefficients, or other changes to prevent poor fitting. In general, however, while these alterations can improve results over a wider range of samples, they do not change the basic concept of unmixing, which is the quantitation of the amounts and locations of the various spectral signatures in spectral image data. While the foregoing appears to be complex, the actual application of unmixing is very straightforward and, in fact, all the complexity can be hidden from the end-user if so desired.

Visualizing Unmixing Results Figure 14.19.1 shows unmixing results from a spectral image acquired from a liver section that had been labeled with Alexa Fluor

Figure 14.19.1 Liver section imaged at 10x using an FITC long-pass emission filter cube from 500 to 720 nm in 10-nm steps. (A) RGB representation of the spectral imaging dataset; (B-D) unmixed images of the tissue autofluorescence, Alexa Fluor 488 and Cy3, respectively; (E) pseudo-color composite image of the combined unmixing results, with autofluorescence in gray, Alexa Fluor 488 in green, and Cy3 in red. For the color version of this figure go to http://www.currentprotocols.com.

14.19.4 Supplement 84

Current Protocols in Molecular Biology

488 for peroxisomes, Cy3 for SE1, Cy5 for actin, and Hoechst for nuclei (Fluor-O-4 liver section, Chroma Technology Corporation) using a FITC/Alexa Fluor 488 long-pass emission filter cube. Fig. 14.19.1A is a representation of the spectral image that uses the eye’s color response to convert the spectral information at each pixel into a red-green-blue color value that can be displayed in an image. This RGB representation is equivalent to what the eye would see looking down the microscope’s eyepieces at the sample. It is clear from Fig. 14.19.1A that there is significant autofluorescence from the liver tissue in the sample. Using a spectral library that consists of spectra for liver autofluorescence, Alexa Fluor 488 and Cy3 give the unmixed images seen in Fig. 14.19.1, panels B to D, respectively. The Hoechst and Cy5 in the sample cannot be detected using an FITC long-pass filter cube, and are therefore not included in the spectral library for unmixing this sample. These unmixed images are the quantitative measures of how much of each fluorophore was present in each pixel. The most common method of visualizing these unmixed images together is to assemble them into a “composite image.” For this, each unmixed image is assigned a color—in

this case, white for autofluorescence, green for Alexa Fluor 488, and red for Cy3—and then the three colored images are combined together into a pseudo-colored composite image (Fig. 14.19.1E). This composite image shows the relative locations and intensities of the fluorophores. If the autofluorescence signals were displayed in black instead of white, then this component would no longer be visible, and the result would reveal what the sample would have looked like if no autofluorescence had been present. One of the potential benefits of unmixing is the sometimes dramatically increased contrast in images in which a fluorophore signal has been enhanced by removing the effects of autofluorescence and/or cross-talk between fluorophores. Figure 14.19.2A shows a detail from Figure 14.19.1A in which the green Alexa Fluor 488 and the red Cy3 are barely visible above the ubiquitous liver autofluorescence. Fig. 14.19.2B shows the matching detail from the composite image, where close examination shows that the unmixing results are correct. The increase in contrast provided by unmixing is best shown by comparing the unmixed results for a particular fluorophore to the monochrome image acquired at the peak emission wavelength of that

Figure 14.19.2 Zoomed images from liver section in Figure 14.19.1. (A) RGB representation; (B) composite image; (C) unmixed Alexa Fluor 488 image; (D) monochrome image captured at peak of Alexa Fluor 488 emission (530 nm). For the color version of this figure go to http://www.currentprotocols.com.

In Situ Hybridization and Immunohistochemistry

14.19.5 Current Protocols in Molecular Biology

Supplement 84

Visualizing Microscopy-Based Spectral Imaging Data

fluorophore. Figure 14.19.2, panels C and D, show the Alexa Fluor 488 unmixed image and monochrome image at 515 nm, respectively. Although it is barely possible to see the green Alexa Fluor 488 signal in the RGB image (see Fig. 14.19.2A), it is impossible to differentiate between tissue autofluorescence and Alexa Fluor 488 in Fig. 14.19.2D. Given that the majority of microscopes on which tissue samples are imaged utilize monochrome band-pass emission filters, it is clear that spectral unmixing offers dramatic improvements in signal-to-noise and image contrast. In addition, spectral unmixing raises confidence levels about what is and is not a signal of interest. A critical and often overlooked factor in obtaining correct results from linear unmixing is the use of accurate spectral measurements for each of the species expected to be in the sample. These measured, or, if necessary, measured and then corrected, spectra are what should be used for unmixing; such a set of spectra may be referred to as a spectral “library.” If the library being used for unmixing is incorrect or inappropriate for the sample being unmixed, then poor results will be obtained. There are numerous automated methods for determining spectral libraries, and the software for most commercial systems comes with some form of this, and there have been numerous publications (Eismann and Hardie, 2004; Mu˜noz-Barrutia et al., 2007; Miao and Szu, 2007). However, the best method for determining spectral libraries for multi-labeled samples is to use a series of control samples, each of which has been labeled with just one fluorophore or chromogen. For chromogenic samples, which rarely contain colors other than the chromogens used, the singly labeled control slides are usually sufficient. For fluorescence samples, this series of controls should include a sample that has not been labeled with any fluorophore at all, to provide a pure spectral signature of autofluorescence. If the fluorescence of the labeled fluorophore is bright enough, the spectrum of that fluorophore can be obtained directly from that slide. However, as is often the case, particularly for tissue sections, which usually contain significant autofluorescence, it is insufficient to simply take a spectral sample from the singly labeled slide, as that spectrum will contain contributions from both the fluorophore of interest and the autofluorescence. Using this “mixed” spectrum as a part of a spectral library for unmixing will not provide accurate unmixing. For these types of situations,

a “pure” spectral calculation is necessary, in which an appropriate amount of a pure autofluorescence spectrum is subtracted from the mixed-data spectrum, giving the “pure” spectrum, Spurified , of the fluorophore of interest, as in the following equation: S purified = S mixed − (a × SautoFL ) + offset

Here, a is variable which is a scalar multiplier of the autofluorescence spectrum, SautoFL , and “offset” is an offset. There are various means of determining the correct values for the variables a and “offset” in the above equation, including manual adjustment or some kind of minimization algorithm. Further information can be found elsewhere (Mansfield et al., 2005; Bouchard et al., 2007). In the end, the use of the correct spectral libraries will provide quantitative unmixing into the individual unmixed images. The best means of testing whether a given spectral library is appropriate for a given sample type is to unmix a series of known controls and check whether the unmixing results are correct.

Simulated Bright-Field and Simulated Fluorescence Images Not only fluorescence spectral images can be unmixed. Multilabeled chromogenic samples with spectral imaging data can be collected in bright-field microscopy with appropriate spectral imaging hardware. The unmixing and visualization of the unmixing results is very similar to the procedures described above for fluorescence images. Figure 14.19.3A shows an RGB representation of a spectral image acquired from a sample that had been chromogenically stained for estrogen receptor (ER, with DAB) and progesterone receptor (PR, with Fast Red), with nuclei counterstained with hematoxylin. The unmixed images from this sample can be seen in Figure 14.19.3, panels B to D. These unmixed images can be displayed in a pseudo-color bright-field composite, with the ER, PR, and hematoxylin displayed in red, green, and blue, respectively, against a white background. However, this means of visualizing unmixed images can be less than optimal for highly co-localized markers, as the combination of red, green, and blue signals in a single nucleus can be difficult to discern by eye. For samples like this, a simulated fluorescence display may be more useful. One interesting but subtle feature of simulated displays like this is that, when inverting each layer in the composite image, the color for the

14.19.6 Supplement 84

Current Protocols in Molecular Biology

display of that layer is kept so that, for instance, green intensity on a white background becomes green intensity on a black background, except that white and black (typically reserved for autofluorescence) switch their colors, with black on a white background becoming white on a black background. Figure 14.19.3, panels E and F, show simulated fluorescence images of the unmixing results from the dataset in Fig. 14.19.3A. In Fig. 14.19.3E, all three unmixed images are displayed, with the hematoxylin shown DAPIlike in blue, the ER in red, and the PR in green. The relative distributions of the chromogens can be seen readily. As composite images are composed of the unmixed images, they can be displayed, or not displayed, as desired. Fig. 14.19.3F shows only the ER images (in red) and PR images (in green), without displaying the hematoxylin. The additive color combination of red plus green is yellow, so

co-localized pixels appear as yellow, yellowgreen, or greenish-yellow. Similarly, fluorescence spectral imaging results can be displayed in simulated bright-field composite images. Figure 14.19.4 shows the results from a dual-color immunofluorescence staining of a section of invasive ductal carcinoma stained for CD44v6 (QDot 655) and CD24 (QDot 605), and counterstained with DAPI; this combination of markers may help identify a cancer stem cell (CSC) population. The determination of the protein expression phenotype of CSCs is an important potential application of multi-color IHC and spectral imaging, as serial section analysis is insufficient to determine if a particular cell is a CSC (Ailles and Weissman, 2007). Fig. 14.19.3A shows the RGB representation of the data set. The distribution of the two quantum dots is difficult to see above the strong tissue autofluorescence. The DAPI appears green because

Figure 14.19.3 Invasive ductal carcinoma section stained for estrogen receptor (ER, DAB) and progesterone receptor (PR, Fast Red), and counterstained with hematoxylin. (A) RGB representation of data set; (B-D) unmixed images corresponding to hematoxylin, ER, and PR, respectively; (E) simulated fluorescence composite with hematoxylin in blue, ER in red, and PR in green; (F) simulated fluorescence composite showing only ER (red) and PR (green). For the color version of this figure go to http://www.currentprotocols.com.

In Situ Hybridization and Immunohistochemistry

14.19.7 Current Protocols in Molecular Biology

Supplement 84

Figure 14.19.4 Intraductal carcinoma sample stained for CD44v6 (QDot 655) and CD24 (QDot 605). (A) RGB representation of spectral data; (B-C) unmixed images of CD44v6 and CD24, respectively; (D) pseudo-color composite with DAPI in blue, CD44v6 in red and CD24 in green; (E) simulated brightfield composite with DAPI in blue, CD44v6 in red, and CD24 in green. For the color version of this figure go to http://www.currentprotocols.com.

of the acquisition range of the spectral image (480 to 720 nm). Figure 14.19.4, panels B and C, show the unmixed images for each of the immunomarkers. Figure 14.19.4D shows the standard fluorescence composite image, which provides a great deal of information about co-localization and the cell phenotypes. However, for pathologists who are much more used to viewing bright-field, chromogenically stained samples, the simulated bright-field image in Fig. 14.19.4E may be preferable.

Co-Localization Calculations and Visualizations

Visualizing Microscopy-Based Spectral Imaging Data

One of the major goals of multi-label staining, whether in bright-field or fluorescence, is determining the expression profile, or phenotype, of the cells—i.e., which proteins are being expressed in a single cell or cell compartment. For this purpose, the ability to interactively turn on and off layers in a composite image and to recolor them as desired can be very helpful. Figure 14.19.5A shows an RGB representation of a section of the germinal center of a tonsil that had been chromogenically stained for Ki-67 (Fast Red) for cells in G1 through M phase of the cell cycle, phospho-histone H3 (pHH3, DAB) for the fraction of cells in mi-

tosis, cleaved caspase 3 (CC3, gray) for apoptotic cells, and hematoxylin (blue) as a counterstain, giving a total of four nuclear markers and four unmixed images (not shown), one for each marker. Figure 14.19.5, panels B to D, shows simulated fluorescence composite images with the layers displayed one at a time, with Ki-67 in red, pHH3 in green, and CC3 in blue. An interactive means of displaying these layers, not possible in a publication, can show which cells contain which of the markers, allowing a visual assessment of co-localization. Often a simple visual estimate of co-localization is insufficient for a particular experiment, and more quantitative analyses are desirable. There are several means of estimating appropriate statistics, the simplest of which is to perform pixel-based co-localization calculations on thresholded unmixed images. Figure 14.19.6 shows the results of this thresholding for the tonsil sample in Figure 14.19.5. Fig. 14.19.6A shows the unmixed hematoxylin image, which forms the basis of the co-localization calculations. It should be noted that the unmixing shows an even hematoxylin staining regardless of the presence or absence of other stains. Figure 14.19.6B shows the thresholded hematoxylin image superimposed on the RGB image of Fig. 14.19.5A.

14.19.8 Supplement 84

Current Protocols in Molecular Biology

Figure 14.19.5 Three nuclear immunohistochemical markers (Ki67, pHH3, CC3) and a nuclear counterstain (hematoxylin) distinguished using chromogens in bright-field in the germinal center of a tonsil. (A) Shows the RGB representation of the dataset; (B) simulated fluorescence display of Ki-67 (red); (C) simulated fluorescence display of Ki-67 (red) and pHH3 (green); (D) simulated fluorescence display of Ki-67 (red), pHH3 (green), and CC3 (blue). For the color version of this figure go to http://www.currentprotocols.com.

Any pixel whose hematoxylin value is above an adjustable threshold is shown in blue. Fig. 14.19.6C shows the thresholded Ki-67 (red) and pHH3 (green) images superimposed on the RGB image. The red and green thresholded layers were added sequentially, with the red being drawn first, followed by the green on top, so one is unable to determine directly from the single image in Figure 14.19.6B whether the pHH3-positive (green) nuclei are also Ki67-positive (red). To get around the limitation of drawing order on determining whether a given pixel contains two or more markers, a co-localization determination can be performed. This is done using a Boolean AND operator (the intersection of two sets) to determine which pixels are positive for each of the two (or more) markers. If one does this for Ki-67 and pHH3 and displays the co-localized pixels in yellow superimposed on Fig. 14.19.6C, one obtains the

image seen in Figure 14.19.6D. All of the yellow pixels contain both Ki-67 and pHH3. The pixels that are only Ki-67 positive are still seen in red, and the pixels that are pHH3 positive are still shown in green. The Boolean AND operator and display of co-localized pixels in some appropriate color is a good means of making a visual assessment of co-localization. However, often one requires a more quantitative determination of the degree of co-localization. In that case, a full matrix of the percentage of pixels that contain two markers can be determined. The results for the sample in Figures 14.19.5 and 14.19.6 can be seen in Table 14.19.1. The table shows, for instance, that 83% of the pixels that contain pHH3 also contain Ki-67 across the entire image, and that only 3% of the CC3 pixels contain pHH3. In addition, the percent positivity for each marker is shown in the bottom row. This value is defined as the percentage

In Situ Hybridization and Immunohistochemistry

14.19.9 Current Protocols in Molecular Biology

Supplement 84

Figure 14.19.6 Co-localization analysis of Ki-67 and pHH3 in tonsil sample from Fig. 14.19.5. (A) unmixed hematoxylin image (blue on white); (B) thresholded hematoxylin image with positive hematoxylin pixels in dark blue superimposed on RGB representation (Fig. 14.19.5A); (C) thresholded Ki-67 (red) and pHH3 (green) images superimposed on RGB representation; (D) the same image as 6C but with co-localized pixels (those containing both Ki-67 and pHH3) shown in yellow. For the color version of this figure go to http://www.currentprotocols.com. Table 14.19.1 Percentage Co-Localization of Pixels of Marker A in Marker B

Marker A Marker B

Visualizing Microscopy-Based Spectral Imaging Data

Ki-67

pHH3

CC3

Ki-67

—

16%