Abstract
Quantitative analysis of digitized IHC-stained tissue sections is increasingly used in research studies and clinical practice. Accurate quantification of IHC staining, however, is often complicated by conventional tissue counterstains caused by the color convolution of the IHC chromogen and the counterstain. To overcome this issue, we implemented a new counterstain, Acid Blue 129, which provides homogeneous tissue background staining. Furthermore, we combined this counterstaining technique with a simple, robust, fully automated image segmentation algorithm, which takes advantage of the high degree of color separation between the 3-amino-9-ethyl-carbazole (AEC) chromogen and the Acid Blue 129 counterstain. Rigorous validation of the automated technique against manual segmentation data, using Ki-67 IHC sections from rat C6 glioma and β-amyloid IHC sections from transgenic mice with amyloid precursor protein (APP) mutations, has shown the automated method to produce highly accurate results compared with ground truth estimates based on the manually segmented images. The synergistic combination of the novel tissue counterstaining and image segmentation techniques described in this study will allow for accurate, reproducible, and efficient quantitative IHC studies for a wide range of antibodies and tissues. (J Histochem Cytochem 56:873–880, 2008)
Keywords: immunohistochemistry, quantitative analysis, image segmentation, tissue counterstaining
Image analysis of digitized IHC-stained tissue sections provides a powerful tool for quantification of protein expression. The use of quantitative IHC (qIHC) has increased in recent years for numerous applications, including diagnostic and prognostic determinations in the clinical setting, and correlation with complementary quantitative measures, such as real-time PCR, in research laboratories. Although many investigators perform qIHC analysis on chromogen-stained tissue sections without the use of a counterstain, tissue counterstaining presents a number of practical advantages. Among these advantages are orientation of tissue on the slide, assessment of tissue morphology, specific determination of positively and negatively staining structures and cells, and intensification of chromogen staining for improved visualization. For these reasons, IHC sections are routinely counterstained in most anatomical pathology laboratories. Furthermore, the recent advent of commercially available, ultra-high-resolution, digital slide scanners has heightened interest in whole slide and whole tissue qIHC measures for both clinical and research applications. Accurate three-dimensional reconstruction of qIHC sections also requires counterstained tissue to allow for proper alignment of the two-dimensional sections comprising the three-dimensional volume (Brey et al. 2002; Chakravarty et al. 2007).
Counterstaining of IHC-stained tissue, however, complicates segmentation of the digital IHC image into chromogen-positive and chromogen-negative pixels caused by the color convolution of the chromogen and counterstain. A number of methods have been developed to allow for semi- or fully automated segmentation of IHC images. These techniques have been thoroughly reviewed and rigorously compared by Brey et al. (2003). Hematoxylin is commonly used as a counterstain for IHC, primarily because of the relatively high level of contrast produced between the brown or red chromogens, such as DAB and 3-amino-9-ethyl-carbazole (AEC), and the blue color of hematoxylin. However, this particular counterstain introduces complications for segmentation algorithms. Brey et al. (2003) pointed to the fact that nuclear IHC staining confounds the results of all evaluated techniques and that nuclei with mixed high-intensity hematoxylin and chromogen staining will result in inconsistent results. Furthermore, variability in the intensity of nuclear staining leads to heterogeneous nuclear chromogen staining, which is independent of the actual amount of antigen present, thereby effectively precluding accurate quantitative intensity and absolute/relative concentration measures. Because nuclear chromogen-staining represents a significant proportion of IHC studies, an improved strategy for nuclear qIHC analysis is warranted.
One strategy for improving the accuracy of image segmentation and classification is to increase the contrast between the component-of-interest and the background tissue and use a tailored segmentation algorithm based on the resultant image contrast. Such a strategy has been successfully used in other areas of medical imaging, including analysis of multiple sclerosis (MS) lesions on MRI scans (Bedell et al. 1997). Translating this approach to digitized IHC images involves improving contrast between chromogen and counterstain followed by the application of a simple segmentation algorithm, with few explicitly specified constraints, thereby maximizing the applicability of this technique to multiple, different antigens and tissues.
In this study, we describe the development of a new tissue counterstain and associated slide preparation techniques that results in homogeneous background staining of tissue, as well as a straightforward, robust, automated image segmentation and classification algorithm that takes advantage of the high degree of separation between the light blue color of the counterstain and the dark-red color of the AEC chromogen. The accuracy of this technique has been evaluated by comparison of the results of the automated algorithm with manual classification of positively stained nuclei on rat C6 glioma tissue IHC stained for the nuclear cell cycle marker Ki-67, as well as on brain tissue from transgenic mice with amyloid precursor protein (APP) mutations, which has been IHC stained for β-amyloid.
Materials and Methods
Animal Models and Tissue Preparation
Rat C6 glioma cells were purchased from American Type Culture Collection (Rockville, MD) and grown in DMEM supplemented with 10% FBS, 125 U/ml penicillin G, 125 μg/ml streptomycin sulfate, and 2.2 μg/ml amphotericin B (Fungizone). All culture reagents were obtained from Gibco BRL (Invitrogen; Burlington, Ontario, Canada). Cultures were grown in monolayers and maintained at 37C in a humidified atmosphere of 5% CO2. On reaching confluency, spheroids were prepared using the hanging drop method previously described by Del Duca et al. (2004). Briefly, 20-μl drops of DMEM containing 15,000 cells each were suspended from the lids of culture dishes, and the resulting aggregates were transferred to culture dishes base-coated with agar after 72 hr. The resulting spheroids were adequate for in vivo implantation after 48 hr of incubation on agar.
All animal experiments were conducted in accordance with the guidelines of the Canadian Council on Animal Care and the Montreal Neurological Institute and McGill University Institutional Animal Care and Use Committees. Five male Sprague-Dawley rats (250–300 g; Charles River Canada, St. Constant, Quebec, Canada) were anesthetized with 50 mg/kg ketamine and10 mg/kg xylazine. The right cortical surface in the parietal-occipital region was exposed by craniectomy using a high-powered drill (DREMEL; Racine, WI), and the underlying dura and its vessels were carefully removed under a surgical microscope. A piece of the cortex was removed to expose the underlying white matter, and a single spheroid was placed into the surgical defect. The craniectomy was covered with bone wax (Ethicon; Peterborough, Canada), and the overlying skin was sutured. After recovery from anesthesia, the animals were fed and had access to water ad libitum. After 16–18 days, animals were sacrificed, and brains were removed and immediately immersed in 10% neutral-buffered formalin.
Tissue with β-amyloid deposition was obtained from the brains of four 18-month-old transgenic APPSw/Ind line J20 mice (Mucke et al. 2000). All mice were deeply anesthetized with an overdose of ketamine/xylazine and perfused with 10% neutral-buffered formalin by intracardiac puncture. After perfusion, the brains were extracted and immersion-fixed in 10% neutral-buffered formalin.
IHC
Brains were fixed for 48 hr, dehydrated through graded alcohols and xylene, and embedded in paraffin wax. A single 5-μm section was cut through the center of each tumor for the rats or through the hippocampus for the transgenic mice using a rotary microtome, and each section was separately mounted on positively charged microscope slides (Fisher Scientific; Ottawa, Ontario, Canada). C6 glioma tissue sections underwent heat-induced antigen-retrieval in boiling 10 mM citrate buffer (pH = 6.1) at elevated pressure (120 psi) (Qi et al. 2006). APP mouse brain sections were antigen-retrieved for 5 min at room temperature in 80% formic acid (pH = 0.5) followed by washing in TBS (Kitamoto et al. 1987). All IHC studies were performed using reagents obtained from Lab Vision (Fremont, CA) and performed on a Lab Vision Autostainer 360. Rat C6 glioma tissue was stained with monoclonal Ki-67 (1:200, 120 min) followed by visualization with the Ultravision LP detection system and AEC chromogen, whereas the transgenic APP mouse tissue was stained with rabbit anti-β-amyloid (1:30, 60 min), followed by amplification with anti-rabbit biotin-streptavidin-horseradish peroxidase (HRP) and visualization with AEC chromogen.
Tissue Counterstaining and Slide Preparation
IHC-stained slides were counterstained for 1 min in 0.1% Acid Blue 129 (Sigma-Aldrich Canada; Oakville, Ontario, Canada), rinsed for 30 sec in sodium acetate-acetic acid buffer (pH = 3.6), and air-dried. We encountered difficulties associated with bleeding of the anionic Acid Blue 129 stain into commercially available aqueous mounting media. Although this counterstain was found to be stable in xylene- or toluene-based mounting media, which could be used for DAB-stained IHC sections, the AEC chromogen required the use of an aqueous mounting media. As such, we developed and used a specially formulated mounting media, consisting of 40% (w/v) PEG 8000 (Sigma-Aldrich) and 20% glycerol (v/v) in sodium acetate-acetic acid buffer (pH = 3.6) to minimize bleeding of the dye from the tissue. Although this media prevented significant bleeding for >1 week, sealing the edges of the coverslips, with nail polish for example, allowed for long-term slide storage. Given that all of our slides were digitized immediately after staining, long-term storage was not an issue. Alternatively, the Acid Blue 129 can be removed and restained at a later time for rereview of slides, or slides can be simply recounterstained with hematoxylin for long-term storage. In this study, we performed the latter procedure to directly compare Acid Blue 129 and hematoxylin counterstaining.
After slide digitization (described in the following section), the coverslips were removed by soaking slides in distilled water for 5–10 min. The Acid Blue 129 counterstain was readily removed by soaking slides in 2 mM aqueous ammonia for 5–10 min. This technique completely removed the Acid Blue 129 while preserving the AEC chromogen. The slides were subsequently counterstained with hematoxylin for 1 min and mounted in permanent aqueous mounting media (Aquatex; EMD Chemicals, Gibbstown, NJ). These hematoxylin-counterstained sections were digitized for comparison to the Acid Blue 129 counterstaining.
Imaging
Slides were digitized using a Zeiss MIRAX Scan ultra-high-resolution, automated slide scanner (Carl Zeiss Canada; Toronto, Ontario, Canada). This system uses a “compensation image” for automated white balancing and normalization to the dynamic range of the system. For each slide, the system averages RGB intensity information from 10 empty (i.e., background) fields of view (FOVs) to produce a “compensation FOV” or “compensation image.” Using this information, the pixels in each FOV captured by the system are normalized by the respective color component from the compensation image, on a pixel-by-pixel basis, and scaled to an intensity range between 0 and 255.
Regions of interest (ROIs) were selected from each of the slides for comparison of the automated segmentation algorithm with manual segmentation. Specifically, cellular regions and regions adjacent to necrotic areas were identified on each Ki-67 IHC slide, whereas regions containing both parenchymal β-amyloid plaques and vascular β-amyloid deposits were identified on each β-amyloid IHC slide. ROIs were captured at ×400 effective magnification and stored in the bitmap (BMP) image file format.
Manual Segmentation
To validate the automated segmentation technique against a gold standard, two estimates of the “ground truth” were obtained using manual segmentations from two independent raters. Different ROIs from Ki-67 (cellular areas, n=5; perinecrotic areas, n=5) and β-amyloid (n=6) IHC sections counterstained with Acid Blue 129 were manually segmented once by each rater. In each ROI, the raters labeled only those pixels that they deemed to represent true positive staining using the “paintbrush” tools in Photoshop (Adobe; San Jose, CA).
Automated Image Segmentation Algorithm
The automated algorithm was based on the high degree of color separation between the red-brown AEC chromogen and the light-blue counterstain. The algorithm first eliminated all white, background (non-tissue) pixels from the analysis. It then calculated the ratio of the blue-to-red channel (B/R ratio) for each pixel. For the Acid Blue 129-counterstain, pixels were classified as “chromogen-positive” based on a predefined criterion, specifically B/R ratio < 1.4 and Green Channel < 140 (intensity range = 0–255). The threshold on the intensity of the green channel was used to minimize false positives resulting from regions of bright counterstain. To directly compare the performance of the algorithm using Acid Blue 129–counterstained slides to the corresponding hematoxylin-counterstained slides, a separate criterion was derived for the hematoxylin counterstain, namely B/R ratio < 1.25 and Green Channel < 110 (intensity range = 0–255). The lower values for the ratio and Green Channel parameters for the hematoxylin, in comparison to the Acid Blue 129, reflect the lesser degree of separation between the colors of the AEC chromogen and the counterstain. The criteria for both counterstains were optimized on a trial set of images, which were independent from the images used for the manual and automated segmentations. Once the threshold criterion was fixed, the process of counting positive and negative pixels was completely automated for the analysis of all images. The algorithm was implemented using ImageJ (Abramoff et al. 2004).
Analysis
The automated segmentation technique was evaluated against the ground truth estimates provided by the manual segmentations using three different evaluation metrics, namely the κ overlap, the true positives, and the false positives metrics. To calculate these metrics, we used set theory to define three specific regions, specifically 1 (the region defined only by the ground truth, i.e., manual estimate), 2 (the region defined by the ground truth estimate and the competing technique, i.e., automated technique), and 3 (the region defined only by the competing technique), as depicted in Figure 1. The number of pixels in each of these regions was defined as P1, P2, and P3, respectively. Using these variables the following analyses were performed:
Figure 1.
Illustration of regions defined for set theory-based analysis of image segmentation data. Region 1 represents the ground truth (i.e., manual estimate), Region 2 represents the ground truth estimate and the competing (i.e., automated) technique, and Region 3 represents the competing technique.
The κ Metric
The κ overlap metric was used to assess the level of agreement between the competing technique and the ground truth estimate, where
 |
Typically, κ > 0.7 is deemed to be an acceptable threshold in image segmentation and classification (Chakravarty et al. 2005). The κ metric is known to reward methods for high levels of pixel agreement while penalizing methods for high levels of pixel disagreement.
Percentage of True Positives
The percentage of true positives (TP) metric was calculated as the proportion of pixels from the ground truth estimate, which were labeled by the competing technique, where
 |
Percentage of False Positives
The percentage of false positives (FP) metric was calculated as the percentage of pixels labeled by the segmentation technique, which were not in agreement with the ground truth estimate, where
 |
The quality of the ground truth estimates was first evaluated to assess the level of agreement between the two manual raters. The TP and FP were calculated by alternating each of the manual rater's labels as the ground truth estimate. These values were examined using a multivariate ANOVA (MANOVA) with the raters' values as the main effect and covariates of IHC stain. Statistical differences were determined with two tailed t-tests.
The automated segmentation technique was assessed against the ground truth estimates provided by the manual raters. The results from the manual raters (i.e., single κ value and a pair of TP and FP values per image) were used as competing techniques. A repeated-measures MANOVA was performed to determine whether statistically significant differences existed between the each of the techniques being analyzed, using IHC stain as a covariate. In cases where statistically significant differences were observed, ANOVA and post hoc Tukey-Kramer honestly significantly different (HSD) tests were performed to group the techniques. The grouping of the post hoc Tukey-Kramer HSD identified the source of the significant differences in the MANOVA or ANOVA analyses. For example, methods classified as Group A had no statistical intragroup differences but were significantly better than methods classified as Group B.
Results
Automated Segmentation of Ki-67 and β-amyloid IHC-stained Tissue Sections
The automated algorithm was found to detect AEC chromogen-positive pixels with a high level of sensitivity for both the Ki-67 and β-amyloid IHC-stained tissue sections, which had been counterstained with Acid Blue 129. Representative examples of the results generated by the automated algorithm using Acid Blue 129–counterstained, Ki-67 and β-amyloid IHC-stained tissue sections are shown in Figure 2. Blood was not found to be a confounding factor for the automated segmentation, because the erythrocytes stained dark blue (Figure 2A, arrow), whereas the necrotic regions were pale blue (Figure 2A, arrowhead). As such, the blood had a high blue-to-red ratio and was well separated from the dark-red AEC chromogen.
Figure 2.
Automated segmentation of Acid Blue 129 counterstained, Ki-67 IHC-stained rat C6 glioma tissue sections and β-amyloid IHC-stained transgenic APP mouse brain tissue sections. (A) Ki-67 IHC staining of rat C6 glioma (arrow, erythrocytes; arrowhead, necrosis), (B) segmented Ki-67 IHC image, (C) β-amyloid IHC staining of transgenic APP mouse brain, and (D) segmented β-amyloid IHC image. Bar = 50 μm.
The Acid Blue 129 counterstain provided a greater contrast-to-noise ratio (CNR) than the hematoxylin counterstain, thereby allowing for the use of higher blue-to-red ratio and green intensity level thresholds in the segmentation algorithm to maximize TP while minimizing FP. Furthermore, the homogeneous nuclear and cytoplasmic staining provided by the Acid Blue 129 eliminated the false negatives resulting from the color convolution of the chromogen and dark nuclear hematoxylin staining in nuclei with condensed chromatin (e.g., vascular endothelial cells). Figure 3 shows a side-by-side comparison of the same IHC ROIs counterstained with Acid Blue 129 and hematoxylin. Note that the use of the Acid Blue 129-counterstain reduced the number of false negatives on both the segmented Ki-67 and β-amyloid IHC images.
Figure 3.
Direct comparison of Acid Blue 129 and hematoxylin counterstains. (A) Ki-67 IHC staining of rat C6 glioma with Acid Blue 129 counterstain and (B) automatically segmented image. (C) Corresponding Ki-67 IHC staining of rat C6 glioma with hematoxylin counterstain and (D) automatically segmented image. (E) β-amyloid IHC staining of transgenic APP mouse brain with Acid Blue 129 counterstain and (F) automatically segmented image. (G) Corresponding β-amyloid IHC staining of transgenic APP mouse brain with hematoxylin counterstain and (H) automatically segmented image. Note that the Acid Blue 129 counterstain produces a homogeneous staining pattern, whereas the hematoxylin counterstain strongly delineates the cell nuclei. False negative can be readily identified on the IHC sections counterstained with hematoxylin because of reduced CNR compared with Acid Blue 129 (arrows in A–D), as well as because of color convolution in darkly hematoxylin-stained nuclei (arrowheads in A–D and asterisks in E–H). Bar = 50 μm.
Comparison of Manual and Automated Segmentation Data
The TP and FP results for the manual segmentations of the data (summarized in Table 1) were analyzed to determine the level of agreement between the manual raters. The MANOVA analysis showed no effect of IHC stain (F = 1.0651, df = 1, p<0.3174), but a significant effect of rater (F = 3.0141, df = 1, p<0.0157). The results pooled for each of the stains showed significant inter-rater differences in the β-amyloid stain (t = 3.17, df = 5, p<0.033) but not for the Ki-67 stain (t = −1.17, df = 9, p<0.865). These results indicated high inter-rater variability with respect to the definition of positive β-amyloid staining, but a high level of agreement in the definition of positive Ki-67 staining. FP analysis also showed no effect of IHC (F = 1.0651, df = 1, p<0.3174); however, significant differences between raters were observed (F = 5.865, df = 1, p<0.0353). When these results were pooled across each stain, significant differences between β-amyloid segmentations were also identified (t = 3.17, df = 5, p<0.033), whereas no significant differences for the Ki-67 segmentations were observed (t = −1.09, df = 9, p<0.750). This analysis further confirmed the inter-rater variability in the segmentations of β-amyloid plaques, thereby underscoring the intrinsic limitations of manual segmentation techniques.
Table 1.
TP and FP analysis of manual segmentation data
| Test | IHC | Rater | Mean ± SD | t–value (p) |
|---|---|---|---|---|
| TP | Both | 1 | 82.9 ± 12.1 | |
| 2 | 84.6 ± 13.0 | |||
| TP | β-amyloid | 1 | 74.4 ± 18.0 | 3.17 (0.033)* |
| 2 | 88.2 ± 16.0 | |||
| TP | Ki-67 | 1 | 89.7 ± 10.9 | −1.17 (0.865) |
| 2 | 81.8 ± 15.9 | |||
| FP | Both | 1 | 16.2 ± 13.4 | |
| 2 | 17.2 ± 15.4 | |||
| FP | β-amyloid | 1 | 16.0 ± 14.2 | 3.17 (0.033)* |
| 2 | 22.3 ± 16.5 | |||
| FP | Ki-67 | 1 | 18.2 ± 11.0 | −1.09 (0.750) |
| 2 | 12.0 ± 10.2 |
*Statistically significant.
TP MANOVA effect of rater: (F = 3.0141, df = 1, p<0.0157).*
TP MANOVA effect of IHC: (F = 1.0651, df = 1, p<0.3174).
FP MANOVA effect of rater: (F = 5.865, df = 1, p<0.0353).*
FP MANOVA effect of IHC: (F = 1.0651, df = 1, p<0.3174).
TP, true positive; FP, false positive; Both, β-amyloid and Ki-67; MANOVA, multivariate ANOVA.
The κ overlap metric was used to evaluate the accuracy of the automated technique compared with the ground truth estimate. The results of the ANOVA analysis for the κ metric are shown in Table 2. Although the MANOVA analysis did not show any significant effect of the IHC stain (F = 2.25, df = 1, p<0.1531), it did show a significant effect of the segmentation technique (F = 4.79, df = 2, p<0.025). The Tukey-Kramer HSD post hoc analysis showed that the κ value of the automated segmentation technique had a high level of agreement compared with Rater 1 (Group A) but a lower level of agreement with Rater 2 (Group B), indicating the variability associated with the manual segmentation technique.
Table 2.
Results of ANOVA for κ data
| Test | Technique | Tukey's post hoc HSD group | Mean ± SD |
|---|---|---|---|
| Both IHC stains | Raters | A | 0.835 ± 0.096 |
| Auto vs. rater 1 | A | 0.836 ± 0.109 | |
| Auto vs. rater 2 | B | 0.760 ± 0.066 | |
| β-amyloid | Raters | NA | 0.822 ± 0.145 |
| Auto vs. rater 1 | 0.833 ± 0.034 | ||
| Auto vs. rater 2 | 0.701 ± 0.120 | ||
| Ki-67 | Raters | NA | 0.845 ± 0.031 |
| Auto vs. rater 1 | 0.838 ± 0.097 | ||
| Auto vs. rater 2 | 0.813 ± 0.072 |
*Statistically significant.
MANOVA effect of IHC (F = 2.25, df = 1, p<0.1531).
MANOVA effect of segmentation technique (F = 4.79, df = 2, p<0.025).*
HSD, honestly significant difference; NA, Tukey's post hoc test not applicable; MANOVA, multivariate ANOVA.
The nature of the differences observed between the manual and automated segmentation techniques was further explored. The results of the ANOVA for the TP tests are shown in Table 3. The MANOVA analysis did not show any significant effect of the IHC stain (F = 1.0651, df = 1, p<0.3174). However, this analysis indicated a significant effect of the segmentation technique (F = 5.29, df = 3, p<0.0353). The Tukey-Kramer HSD post hoc analysis showed that the TP value of the automated segmentation technique had a higher level of agreement compared with Rater 1 or Rater 2 (Group A) than the inter-rater comparison.
Table 3.
Results of ANOVA for TP data
| Test | Technique | Tukey's post hoc HSD group | Mean ± SD (%) |
|---|---|---|---|
| Both IHC stains | Rater 1 vs. rater 2 | B | 82.9 ± 12.1 |
| Rater 2 vs. rater 1 | A B | 84.6 ± 13.0 | |
| Auto vs. rater 1 | A | 94.1 ± 5.9 | |
| Auto vs. rater 2 | A | 92.2 ± 13.0 | |
| β-amyloid | Rater 1 vs. rater 2 | NA | 74.4 ± 18.0 |
| Rater 2 vs. rater 1 | 88.2 ± 16.0 | ||
| Auto vs. rater 1 | 95.0 ± 3.5 | ||
| Auto vs. rater 2 | 94.6 ± 15.7 | ||
| Ki-67 | Rater 1 vs. rater 2 | NA | 89.7 ± 10.9 |
| Rater 2 vs. rater 1 | 81.8 ± 15.9 | ||
| Auto vs. rater 1 | 93.2 ± 2.5 | ||
| Auto vs. rater 2 | 90.4 ± 15.7 |
*Statistically significant.
MANOVA effect of IHC (F = 1.0651, df = 1, p<0.3174).
MANOVA effect of segmentation technique (F = 5.29, df = 3, p<0.0353).*
HSD, honestly significant difference; NA, Tukey's post hoc test not applicable; MANOVA, multivariate ANOVA.
The results of the ANOVA for the FP tests are shown in Table 4. In this case, significant differences were observed for both the segmentation technique (MANOVA results: F = 4.75, df = 3, p<0.0049) and the IHC stain (MANOVA results: F = 12.11, df = 1, p<0.0009). Further analysis of the individual IHC stains by ANOVA indicated that there was a significant effect of segmentation technique for the β-amyloid stain (F = 8.40, df = 3, p<0.0004) but not for the Ki-67 stain (F = 1.9734, df = 3, p<0.145). The Tukey-Kramer HSD post hoc analysis for the β-amyloid stain showed a high level of variability between the segmentation techniques.
Table 4.
Results of ANOVA FP data
| Test | Technique | Tukey's post hoc HSD group | Mean ± SD (%) |
|---|---|---|---|
| Both IHC stains | Rater 1 vs. rater 2 | NA | 16.2 ± 13.4 |
| Rater 2 vs. rater 1 | 17.2 ± 15.4 | ||
| Auto vs. rater 1 | 24.5 ± 12.1 | ||
| Auto vs. rater 2 | 28.3 ± 14.8 | ||
| β-amyloid | Rater 1 vs. rater 2 | A | 16.0 ± 14.2 |
| Rater 2 vs. rater 1 | A B | 22.3 ± 16.5 | |
| Auto vs. rater 1 | B C | 25.4 ± 4.3 | |
| Auto vs. rater 2 | C | 32.3 ± 13.0 | |
| Ki-67 | Rater 1 vs. rater 2 | NA | 18.2 ± 11.0 |
| Rater 2 vs. rater 1 | 12.0 ± 10.2 | ||
| Auto vs. rater 1 | 23.9 ± 15.1 | ||
| Auto vs. rater 2 | 22.3 ± 17.0 |
*Statistically significant.
MANOVA effect of IHC (F = 12.11, df = 1, p<0.0009).*
MANOVA effect of segmentation technique (F = 4.75, df = 3, p<0.0049).*
ANOVA β-amyloid, effect of segmentation technique (F = 8.40, df = 3, p<0.0004).*
ANOVA Ki-67, effect of segmentation technique (F = 1.9734, df = 3, p<0.145).
HSD, honestly significant difference; NA, Tukey's post hoc test not applicable; MANOVA, multivariate ANOVA.
Discussion
We successfully developed a robust, fully automated method for the quantification of AEC-stained IHC sections. The Acid Blue 129 stain provided homogeneous tissue counterstaining and a high degree of contrast between tissue and chromogen. The use of this new counterstaining system allowed for the implementation of a simple segmentation algorithm, which takes advantage of the high level of color separation between the dark-red AEC chromogen and the uniform, light-blue Acid Blue 129 counterstain. Once the optimal threshold criteria were determined based on trial images, and fixed for the algorithm, this automated technique provided results comparable to the ground truth manual segmentation results.
When the automated segmentation technique was compared with both raters, as assessed by the κ metric (Table 2), varying degrees of overlap were observed. No statistical differences were observed between the inter-rater κ (Tukey-Kramer HSD post hoc test Group A) and the κ value of the automated segmentation technique compared with Rater 1 (Group A). However, a statistically significant difference was observed when comparing the automated segmentation technique to Rater 2 (Group B). This comparison effectively showed the variability inherent in the manual segmentation. The TP and FP tests were used to further examine the types of differences observed in the data. The results of the TP test established that the automated segmentation technique had a significantly higher degree of similarity to both raters than the similarity observed between individual raters. The results of the FP test, on the other hand, showed a high degree of variability when comparing the automated segmentation technique to both raters. The overall analysis indicated that the automated segmentation technique was as accurate as the manual raters, whereas the automated technique does not suffer from the intrinsic confounds of high inter-rater bias and error.
We determined that the use of the Acid Blue 129 counterstain minimized false positives and false negatives compared with the hematoxylin counterstain. The higher CNR provided by the Acid Blue 129 allowed for a greater detection of true chromogen staining while minimizing false positives in the background tissue. Furthermore, the homogenous staining produced by the Acid Blue 129 eliminated the convolution of dark-blue nuclear staining with chromogen staining, which is often produced with hematoxylin counterstaining and can result in false negatives. The superior performance of the algorithm using Acid Blue 129–counterstained sections compared with hematoxylin-counterstained tissue may be advantageous for IHC studies in which the tissue morphology is dominated by cells with highly condensed nuclear chromatin (e.g., neuroendocrine tumors and lymphomas).
The IHC quantification method used in this study was based on a binary classifier (i.e., chromogen positive or chromogen negative). Significant interest exists, however, in quantification of the absolute or relative level of protein expression (Shi et al. 2005). Although an optimal method of IHC standardization remains an unsolved challenge, quantification of chromogen intensity is further complicated by convolution of the chromogen and counterstain color components (Ruifrok and Johnston 2001). In hematoxylin-counterstained tissues, the intensity of nuclear chromogen staining is function of both the chromogen and variable hematoxylin intensity. The use of Acid Blue 129 should allow for more accurate quantification because it provides a more homogeneous background level of staining compared with hematoxylin.
While Acid Blue 129 is superior to hematoxylin as a counterstain for quantification of IHC sections, it does not provide the same level of morphological detail as hematoxylin. In cases in which both quantification and morphological assessment are desired, Acid Blue 129 can be simply removed by washing with 0.2% aqueous ammonia and the sections subsequently re-counterstained with hematoxylin, as described in the Materials and Methods section. Given that coverslips can be removed from slides prepared with our specially formulated mounting media by soaking in distilled water for ∼10 min and that the chromogen is stable in the aqueous ammonia, this method provides a simple means of obtaining both quantitative and qualitative information.
The recent advent of ultra-high-resolution, whole slide scanners has revolutionized IHC data archiving and quantitative analysis. The fact that all of our slides were digitized using the Zeiss MIRAX Scan system allowed for digital archiving, thereby effectively eliminating any concerns regarding the long-term stability of the Acid Blue 129 counterstain. By digitizing the entire slides, qIHC can be performed over the entire tissue section, rather than being limited to selected ROIs. The large size of these images (e.g., 1010 pixels for a rat brain section) necessitates the use of fully automated qIHC methods. We have recently applied the method described in this study for whole brain qIHC analysis in transgenic mouse models of Alzheimer's disease (Chakravarty et al. 2007).
Given the robustness of the algorithm for the two very different staining patterns produced by the Ki-67 IHC of rat C6 glioma and β-amyloid IHC of transgenic APP mouse brains in this particular study, we anticipate a broad applicability of this technique. Studies are currently underway to rigorously validate this technique using a wide range antibodies, tissues, and chromogens. We believe that the straightforward, inexpensive methods described for the preparation of high-contrast IHC sections and the fully automated segmentation algorithm presented in this study will allow for simple, robust qIHC studies in both research and clinical settings.
Acknowledgments
This work was supported by funds from the Montreal Neurological Institute.
We thank Dr. Edith Hamel for kindly supplying the transgenic APP mouse brains and Kurt Hemmings for assisting in the preparation of the tissue sections and the manual segmentation data.
References
- Abramoff MD, Magelhaes PJ, Ram SJ (2004) Image processing with ImageJ. Biophotonics Int 11:36–42 [Google Scholar]
- Bedell BJ, Narayana PA, Wolinsky JS (1997) A dual approach for minimizing false lesion classifications on magnetic resonance images. Magn Reson Med 37:94–102 [DOI] [PubMed] [Google Scholar]
- Brey EM, King TW, Johnston C, McIntire LV, Reece GP, Patrick CW (2002) A technique for quantitative three-dimensional analysis of microvascular structure. Microvasc Res 63:279–294 [DOI] [PubMed] [Google Scholar]
- Brey EM, Lalani Z, Johnston C, Wong M, McIntire LV, Duke PJ, Patrick CW (2003) Automated selection of DAB-labeled tissue for immunohistochemical quantification. J Histochem Cytochem 51:575–584 [DOI] [PubMed] [Google Scholar]
- Chakravarty MM, Collins DL, Zehntner SP, Hemmings K, Chan C, Zijdenbos AP, Hamel E, et al. (2007) Ex vivo molecular imaging in transgenic mouse models of Alzheimer's disease. Alzheimer's Association International Conference on Prevention of Dementia, Washington, DC, June 9–13, 2007
- Chakravarty MM, Sadikot AF, Germann J, Bertrand G, Collins DL (2005) Anatomical and electrophysiological validation of an atlas for neurosurgical planning. In Duncan J, Gerig G, eds. Eighth International Conference on Medical Image Computing and Computer Assisted Intervention MICCAI 2005, Volume 2 of Lecture Notes in Computer Science. Palm Springs, FL, Springer, 394–401 [DOI] [PubMed]
- Del Duca D, Werbowetski T, Del Maestro RF (2004) Spheroid preparation from hanging drops: characterization of a model of brain tumor invasion. J Neurooncol 67:295–303 [DOI] [PubMed] [Google Scholar]
- Kitamoto T, Ogomori K, Tateishi J, Prusiner SB (1987) Formic acid pretreatment enhances immunostaining of cerebral and systemic amyloids. Lab Invest 57:230–236 [PubMed] [Google Scholar]
- Mucke L, Masliah E, Yu GQ, Mallory M, Rockenstein EM, Tatsuno G, Hu K, et al. (2000) High-level neuronal expression of Aβ1–42 in wild-type human amyloid protein precursor Tg mice: synaptotoxicity without plaque formation. J Neurosci 20:4050–4058 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Qi ZL, Huo X, Xu XL, Zhang B, Du MG, Yang HW, Zheng LK, et al. (2006) Relationship between HPV16/18 E6 and p53, p21WAF1, MDM2, Ki67 and cyclin D1 expression in esophageal squamous cell carcinoma: comparative study by using tissue microarray technology. Exp Oncol 28:235–240 [PubMed] [Google Scholar]
- Ruifrok AC, Johnston DA (2001) Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol 23:291–299 [PubMed] [Google Scholar]
- Shi S-R, Liu C, Perez J, Taylor CR (2005) Protein-embedding technique: a potential approach to standardization of immunohistochemistry for formalin-fixed, paraffin-embedded tissue sections. J Histochem Cytochem 53:1167–1170 [DOI] [PubMed] [Google Scholar]