ImmunoRatio: a publicly available web application for quantitative image analysis of estrogen receptor (ER), progesterone receptor (PR), and Ki-67

Abstract

Introduction: Accurate assessment of estrogen receptor (ER), progesterone receptor (PR), and Ki-67 is essential in the histopathologic diagnostics of breast cancer. Commercially available image analysis systems are usually bundled with dedicated analysis hardware and, to our knowledge, no easily installable, free software for immunostained slide scoring has been described. In this study, we describe a free, Internet-based web application for quantitative image analysis of ER, PR, and Ki-67 immunohistochemistry in breast cancer tissue sections.

Methods: The application, named ImmunoRatio, calculates the percentage of positively stained nuclear area (labeling index) by using a color deconvolution algorithm for separating the staining components (diaminobenzidine and hematoxylin) and adaptive thresholding for nuclear area segmentation. ImmunoRatio was calibrated using cell counts defined visually as the gold standard (training set, n = 50). Validation was done using a separate set of 50 ER, PR, and Ki-67 stained slides (test set, n = 50). In addition, Ki-67 labeling indexes determined by ImmunoRatio were studied for their prognostic value in a retrospective cohort of 123 breast cancer patients.

Results: The labeling indexes by calibrated ImmunoRatio analyses correlated well with those defined visually in the test set (correlation coefficient r = 0.98). Using the median Ki-67 labeling index (20%) as a cutoff, a hazard ratio of 2.2 was obtained in the survival analysis (n = 123, P = 0.01). ImmunoRatio was shown to adapt to various staining protocols, microscope setups, digital camera models, and image acquisition settings. The application can be used directly with web browsers running on modern operating systems (e.g., Microsoft Windows, Linux distributions, and Mac OS). No software downloads or installations are required. ImmunoRatio is open source software, and the web application is publicly accessible on our website.

Conclusions: We anticipate that free web applications, such as ImmunoRatio, will make the quantitative image analysis of ER, PR, and Ki-67 easy and straightforward in the diagnostic assessment of breast cancer specimens.

Figure 1

A flowchart outlining the ImmunoRatio analysis algorithm. Step 1: A RGB color microscope image, an optional blankfield correction image, and thresholding adjustment parameters are received as an input. Step 2: The blankfield image is used to correct uneven illumination and color balance. If a blankfield image is not available, background subtraction is carried out using the Rolling ball algorithm [19]. Step 3: The Color Deconvolution plugin [20] is used to separate the stains into two eight-bit component images: diaminobenzidine (DAB) and hematoxylin (H). Step 4: The components are processed with a mean filter and binarized using adaptive IsoData thresholding [21]. Component-specific threshold adjustments are applied if defined via input parameters. Step 5: The components are processed with a median filter to smooth the thresholding result. Nucleus segmentation is performed on both components by using the Watershed algorithm [22] and small particles are discarded based on their size. For the H component, thin (fibroblastic) cells are identified and discarded using non-round particle removal. Step 6: The H and DAB components are overlaid on the source image. The percentage of DAB-stained nuclear area out of the total nuclear area (the labeling index) is calculated. An (optional) external calibration function is used to correct the ratio percentage. Step 7: The result image consisting of image identification string, the analysis date, the result labeling index, the original image, and a pseudo-colored image showing the staining components is created. A more detailed algorithm flowchart is available on our research group website [15].

Figure 2

An example result of a Ki-67-stained image processed with ImmunoRatio. The result image includes a sample identifier, the analysis date, the labeling index (percentage of positively stained nuclear area), the original image, and a pseudo-colored image showing the segmented staining components.

Figure 3

A screenshot of the ImmunoRatio web application. The Result window is shown as an insert in the right panel. The application is publicly available on our jvsmicroscope.uta.fi website, where users can analyze their images freely.

Figure 4

Scatter plots comparing labeling indexes defined by visual cell counting, non-calibrated ImmunoRatio, and calibrated ImmunoRatio. (a) The calibration was made using a training set of 50 samples, of which 25 were stained for Ki-67, 13 for progesterone receptor (PR), and 12 for estrogen receptor (ER). To achieve linear relation (dotted line), a correction function was defined by fitting a third degree polynomial (solid black line) to the training set. (b) The calibration was validated by using a separate test set of 50 samples (25 stained for Ki-67, 13 for PR, and 12 for ER). The validation test set included two outliers (marked as brown).

Figure 5

The mean labeling index of ImmunoRatio analysis as a function of the number of images included in the averaged result. Five samples stained for progesteron receptor (PR) and five for Ki-67 were tested.

Figure 6

The importance of optimal immunostaining conditions on the accuracy of ImmunoRatio analysis. The red lines outline the nuclei and highlight the segmentation of (a to c) brown and (d to f) blue staining components. (a) Overly dilute primary antibody concentration (Ki-67 MIB-1, 1:400) causes inadequate brown segmentation. (b) Optimal antibody dilution (1:100). (c) Overly strong antibody concentration (1:25) results in excessive cytoplasmic staining and brown segmentation. (d) Overly dilute hematoxylin staining causes inadequate blue segmentation. (e) Optimal hematoxylin dilution. (f) Overly strong hematoxylin causes excessive cytoplasmic staining and blue segmentation.

Figure 7

Breast cancer-specific survival of 123 breast cancer patients according to the Ki-67 labeling index determined with ImmunoRatio. The cut-off was set at median Ki-67 labeling index (20%). Tumors with a high labeling index were associated with poorer breast cancer-specific survival during the follow up of 20 years (hazard ratio = 2.2, P = 0.01).