Overview of Computer-Assisted Quantification Scheme
The workflow of the quantification scheme is outlined in Figure 1. Standard formalin-fixed, paraffin-embedded HER2 IHC stained diagnostic sections of breast tissue are visualized using diaminobenzidine (DAB) and counterstained with hematoxylin. A relevant microscopic field of the tissue is digitally imaged by a certified pathologist. All pixels which are stained with DAB, indicating areas of HER2 protein, are isolated by utilizing a color decomposition algorithm that was previously described [18]. Subsequently, image-based quantification of HER2 staining proceeds through identifying stained membrane regions using a filter-based algorithm (II) and comparison with positive controls (V). Lastly, we provide a gross estimate of the percentage of positive tumor cells (IV) by factoring the percentage of stained area (expressed in pixels) into the feature analysis.
Case selection, processing, and manual grading
The specimens used in these studies consisted of 99 breast cancer cases that had been diagnosed between January 2005 and March 2007 and stored in archives at the Department of Pathology and Laboratory Medicine at Robert Wood Johnson University Hospital, New Brunswick, and N.J. During this period of time it was standard procedure for both assays to be performed. Cases which had received an IHC score of 0 or 1+ were limited in order to enrich the amount of 2+ cases (considered equivocal) to approximately 25% of the data. The purpose of this experimental design was that since 2+ cases are considered equivocal, they represent the patient population that would tend to benefit most from the use of computer aided quantification. Specimens with significant mechanical crushing and sectioning artifacts were omitted from the study to arrive at the final 99 cases. Both invasive ductal and lobular carcinomas were included in the experiments.
Immunohistochemical staining was performed at the Department of Pathology and Laboratory Medicine at Robert Wood Johnson University Hospital utilizing an automated immunostainer, Ventana BenchMark IHC/ISH system (Ventana Medical Systems, Inc. Tucson, AZ). Ventana PATHWAY HER-2 (clone CB-11) was used for the primary detection of c-erbB-2 antigen in sections of formalin fixed, paraffin embedded tissue. A known 3+, FISH positive control was fixed on each slide along with the patient sample in order to analyze both under identical staining conditions.
Manual immunohistochemical grading was performed by a board certified surgical pathologist (P.J.) at Robert Wood Johnson University Hospital according tithe Scoring Guide for the Interpretation of Ventana Pathway HER2 Staining of Breast Carcinomas [19]. Briefly, if no membrane staining was observed the specimen was scored as 0. Faint, partial staining of the membrane was scored as 1+. Weak complete staining of the membrane, >10% of cancer cells was scored as 2+. Intense complete staining of the membrane, >10% of cancer cells, was scored as 3+. A score of 2+ or greater is considered positive according to the manufacturer's instructions.
Specimens were sent to Genzyme Genetics (Westborough, MA) for FISH analysis and the ratio of discrete signals for the HER2 gene and centromere probe for chromosome 17 (CEP17) was reported. A ratio of HER2 gene/CEP17 ≥ 2.3 was considered positive for HER2 overexpression.
Image Selection and Capturing
A board certified pathologist (P.J.) delineated the region of invasive carcinoma, and then a resident pathologist (M.I.) used a robotic microscope to digitally acquire images at 20× magnification within the specified boundaries. The control from each slide was also digitized at 20× magnification. All images are taken with an Olympus AX70 microscope (Olympus America Inc., Melville, NY) equipped with a Prior six-way robotic stage and motorized turret (Prior Scientific, Inc., Rockland, MA) and Olympus DC330 720-line 3-Chip video camera at 1360 × 1024 pixel resolution and a fixed exposure time of 1/600 s. This allows the digital images to be captured through red, green, and blue channels, and thus, span the entire visible range.
Color Decomposition
The DAB (HER2) stained regions were isolated from the hematoxylin counterstain in the digital image of the patient sample and control images through the use of color decomposition (Fig. 2) [18]. Based on a polar transformation in a color space where the hematoxylin and DAB colors are laid out on a super-plane, color decomposition enables the digital image to be separated into the hematoxylin stained image and the DAB stained image of the tissue. In a sense, this algorithm attempts to describe light absorbance from certain color ranges (DAB and Hematoxlyin) without the use of expensive spectral imaging equipments.
Membrane Isolation Algorithm (MIA)
In order to isolate the membrane stained regions, Otsu's method [20] was used to preprocess the DAB image by automatically removing background noise. Subsequently, a rotationally invariant bar filter was used to detect membranes throughout the DAB stained image. The rotationally invariant filter was created using a set of eight Gaussian based bar filters rotated 2π/8 degrees apart (details in Fig. 3). Similar to [21], rotational invariance is achieved by keeping only the maximum response from the convolution of these 8 filters with the DAB image. Various thresholds were then applied to the maximum response from convolution with the bar filters. The thresholds (k = [1,5,7,9,12,15,20]) used to isolate membrane pixels were evaluated based on their ROC performance. In the results below, only k = 15 data is shown. ROC data were calculated for all thresholds k (see Additional file 1). Membrane isolation results on lighter stained specimens are upon inspection visually more satisfying (see Additional File 2). However experiments conducted using lower thresholds for lighter stained images and higher thresholds for darker images, provided no added benefit in terms of increasing AUC (see Additional file 3).
Image Features and HER2 Score
In these experiments, three scoring features based on mean intensity were reported. The first image feature used to calculate a score is M
p
the mean intensity of the patient's stained membrane regions,
(1)
where I is the set of intensities derived from the pixels retained from performing the MIA on p the patient tissue; the bar denotes the mean function.
2. The second image feature is M
n
which is M
p
normalized by the positive control
(2)
M
c
is defined similarly to M
p
except that the calculation is based on c the control tissue .
3. The third feature M
a
adds a coefficient d/N, to M
n
(3)
where N is the total amount of pixels in the image, d is the number of DAB stained pixels after pre-processing. The coefficient d/N is used as an approximation for the percentage of stained cells. Results were also gathered for median-based features; however, results were similar, and therefore not included in this manuscript.
Hardware and Implementation
An Intel Core 2 Duo T7400 (2.16 GHz) computer with 1.0 GB memory was used for processing each image. The color decomposition algorithm was implemented in Java. All other processing was implemented utilizing Matlab (The MathWorks, Inc. Natwick, MA) code. The color decomposition and membrane isolation together took approximately 10–20 seconds per image to complete.
Statistical Analysis
Receiver Operator Characteristic curve analysis was used to compare the accuracy of the manual and automated computerized methods for grading IHC. FISH assays were used as an independent means for assessing IHC scores. The area -under- the -curve (AUC) was calculated for each ROC curve utilizing the trapezoidal rule [22] in order to compare the accuracy of each method as a test for FISH positivity. Since the data are derived from the same patient cases, p-values are calculated using a nonparametric method based on the Mann-Whitney-U Statistic [22]. The McNemar Test for correlated proportions [23, 24]was used to calculate the p-values for differences between sensitivity or specificity.
Institutional Review Board Approval
Institutional Review Board approval was obtained for this retrospective study through protocol (IRB #5381).