KI67 immunohistochemistry quantification in breast carcinoma: A comparison of visual estimation, counting and Immunoratio©

Abstract

Background

Molecular analysis has shown that breast carcinomas can be classified into several intrinsic subtypes, with implications for management and prognosis. In the majority of pathology laboratories molecular analysis of each case is not possible and immunohistochemistry is used for subtyping. This includes analysis of hormone receptors as well as HER-2-neu and Ki67. The methodology for the interpretation of the proliferation index using Ki67 remains an area of uncertainty. We investigated the degree of agreement between different methods of Ki67 interpretation.

Materials and Methods

We analysed 204 breast core biopsies diagnostic of breast carcinoma using visual estimation/eyeballing (EB), Immunoratio©(IR), and counting by two pathologists (CP1 and CP2). The correlation between the different methods and the inter-observer agreement between the two pathologists was assessed. Specific analysis was also done with respect to classification of cases into low Ki67 groups (using Ki67 values less than 14% and less than 20%) since this is critical in classifying tumours into Luminal A and Luminal B subtypes.

Results

Correlation between the different methods was best achieved comparing IR and CP1, and worst comparing CP1 and EB. Correlation was better when considering inter-observer variability (CP1 vs CP2). Comparing the number of cases classified as low Ki67 (less than 14% and less than 20%) the Cohen’s Kappa statistic varied from κ = 0.267 to κ = 0.814 with different methods. When limiting the analysis to cases with a Ki67 of 10–25% according to any method, there was greater disagreement.

Conclusion

At the higher and lower Ki67 levels, the correlation between the methods of assessment was acceptable, however, at levels close to the cut-off values for Lumial A versus Luminal B, several patients would be differently classified by the different methods and therefore potentially receive suboptimal management.

Introduction

Microarray-based molecular analysis of breast carcinomas has shown distinct sub-groups with a spectrum of molecular and biological characteristics. (1) Perou and Sorlie in 2000 used a hierarchical method of clustering gene expression to divide breast cancers into subtypes that correlate with their morphologic and clinical features. (2) They defined five groups: luminal A, luminal B, normal breast-like, HER2-enriched and basal-like. Subsequently, other subtypes have been described. Similar results have been found with genomic analysis (3) and with multiparameter tests including the PAM50 (1) and MammaPrint©.(4) This has led to a more personalised approach to therapy. (5) Employing molecular methods routinely is often not viable in view of the financial cost and time constraints, and the St. Gallen conference of 2013 advised the use of immunohistochemical stains for classification of breast cancers into intrinsic subtypes. (6) Analysis of these stains has been used to serve as a surrogate assessment as follows: (7)

1. Luminal breast cancer

1. Luminal A: ER and PR positive (PR > 20%), HER2 negative and low Ki67.

2. Luminal B: ER+ PR+ (>20%), HER2−, Ki67 > 14%

   or ER+, HER2−, Ki67 14–19%, and PR < 20%.

1. HER2 Positive

1. Luminal HER2 Subtype: ER+ (but often not as strongly positive as HER2 negative tumours)

2. HER2+−Enriched subtype: ER−, PR−, HER2+

1. Triple negative.

   Basal-like subtype: ER−, PR−, HER2−, Cytokeratin5+ and/or EGFR+

   Non- classified triple negative subtype: Negative for ER, PR, HER2, Cytokeratin5 and EGFR.

   These subtypes are predictive of different responses to chemotherapeutic agents. (8)

During the investigation of specific antigens for diagnosis of lymphoma, the Ki67 antigen was identified as a proliferation marker by Johannes Gerdes. (9) He demonstrated that Ki67 reflects the growth fraction in neoplasms. (10) Prior to this, indirect methods such as 3H thymidine determination and flow cytometry had been used for this purpose (11), but these could not be used on a routine basis. After developing the process of antibody cloning, MIB1 (Molecular Borstol Index) was produced, which was subsequently used to determine the Ki67 index in formalin fixed paraffin embedded tissue (FFPE). (9) Later, the role of the Ki67 protein was further described: to keep the condensed chromosomes dispersed in the cell’s cytoplasm during mitosis. (12) The Ki67 antigen is a nuclear non-histone protein expressed in significant quantities throughout the cell cycle, other than in the quiescent phase and early cell cycle. Anti-Ki67 labelling has been validated as a useful proliferation marker in various neoplasms including haematolymphoid neoplasms, astrocytoma, oligodendroglioma, colon carcinoma and breast carcinoma. (13, 14, 15)

Many, but not all researchers have demonstrated Ki67’s predictive and prognostic value. (14,16) It indicates the likely response to chemotherapy – a better response with a high Ki67. (17) Patients with luminal A breast cancers have been found to benefit most from hormone therapy without chemotherapy, whereas those with luminal B tumours respond better to chemotherapy. (5) Although the use of the Ki67 index in treatment decisions is not universally agreed on, many experts support the individualisation of breast cancer therapy, especially with regards to the use of neo-adjuvant therapy. (6,17,18,19) It has been demonstrated that a large number of patients are likely to be receiving chemotherapy unnecessarily (19). Various methods have been employed to predict the response of individual patients to different therapeutic approaches, but the proliferation status plays an important role. The St. Gallen meeting of 2015 highlighted the dilemma of resource-poor regions of the world where the use of expensive tests and expensive treatments must be rationalised. (18)

There is widespread controversy regarding interpretation of Ki67. It has been found to have unacceptable inter-observer and inter-laboratory variability (20) and some suggest that routine reporting of Ki67 should not be done in breast carcinoma in view of its lack of reliability. (21) Many variables impact the interpretation of Ki67 immunostaining. Various methods are used for assessment of percentage of Ki67 positivity, including “eyeballing” (visual scanning and estimating the percentage staining), formally counting the cells, and image analysis methods. (22) It is uncertain whether the highest staining areas (“hot spots”), average or low staining areas should be assessed, or multiple areas and the mean calculated. Proponents of the first method argue that the “hot spots” are the areas representing the tumour component most likely to metastasize and therefore the most important component prognostically. (22) Assessment of the tumour invasive edge has been recommended. (23) If manual counting of positive cells is done, there is a variable approach to the number of cells counted. The International Ki67 in Breast Cancer Working Group recommended a minimum of 500 cells, and ideally 1000 cells to be counted. (23) One study showed no significant difference between counting 300, 510 or 1020 cells using the standard method, but lower figures counting 1200 cells than 510 cells if using the “hot spot” or advancing edge approach. (24) The “cut-off” value in assigning tumours to the luminal A or luminal B groups is not universally agreed upon. The ideal value has been shown to vary from one institution to another, ranging from 3,5% to 34%. (16) The St. Gallen International Expert Consensus in 2011 recommended a cut-off of 14%. (6) At the St. Gallen conference of 2015, most experts favoured a value of 20–29%. (18) There are also different thresholds obtained if studies are conducted using different clinical end-points and sub-sets of patients. (25) It is recommended that the Ki67 cut-off value should be determined separately for each laboratory. (19)

Image analysis software is usually not freely available and requires virtual microscopy scanning. ImmunoRatio©(IR) is a digital image analysis system which is accessed via a web browser. It is free and no additional equipment or software is required. (26) The website includes a Camera adjustment wizard to set the contrast and brightness of the images. The analysis has been found to be optimal using three images taken with a 20× objective and averaging the results. (26)

Aim

The aim of this study was to compare three different methods of quantifying the proliferation index on Ki67 immunohistochemical stains performed on breast core biopsies, and to assess inter-observer variability in counting by two pathologists.

Methods

This was a retrospective non-interventional study. The original report captured on the National Health Laboratory System (Trakcare) was examined to determine the Ki67 reported on each case, assumed to have been determined using the “eyeballing” (EB) method. Oestrogen receptor positivity in these cases was defined as more than 1% positivity.

The slides stained with the Ki67 immunostain (antibody 30–9, Roche diagnostic, Ventana, USA; using the Ventana Benchmark XT machine) were retrieved from the archives. The Ki67 value was obtained using the counting method (CP1). This was carried out using photomicrographs taken of consecutive fields and 500 cells were counted for each case. Any degree of staining was included. Counting was independently repeated by a second pathologist (CP2), who was blinded to results obtained by CP1. The two pathologists are general pathologists and have similar training and experience.

The slides were then interpreted using the IR software. Photomicrographs of three areas for each case were analysed and a mean value obtained, using a 20× microscopic objective. An Olympus BH-2 microscope with an attached Zeiss AxiocamERc 5s camera was used.

Two hundred and four cases of breast carcinoma diagnosed on core biopsy were obtained from the cases enrolled in the Chris Hani Baragwanath SABC case-control study (PI, I Romieu and funded by the World Cancer Research Fund). Cases with insufficient representation of invasive tumour for counting or IR, where a Ki67 stain was not done or reported, where the patient had been treated with chemotherapy or radiotherapy prior to biopsy, and biopsies with extensive necrosis or inflammation were excluded.

Statistical analysis

The sample size was determined using the formula recommended by Watson et al. (27) A Bland and Altman diagram for each pair was constructed. For each comparison a Lin’s concordance coefficient was determined to assess how well the points lay when plotted with a 45° line drawn though the origin. Comparisons were done of the number of low Ki67 results obtained using cut-off values of 14% and 20% (St. Gallen conference recommended values for 2011 and 2015). A Cohen’s Kappa statistic was calculated for each of these. Since the most important problem is division of the cases into low Ki67 and high Ki67 groups, values within the range of 10 – 25% were also analysed separately.

Results.

Core biopsies were retrieved from female patients newly diagnosed with invasive mammary carcinoma. The patient ages ranged from 27 years to 86 years with a median of 52 years and an interquartile range of 42 – 60 years. All the patients were black Africans.

204 Ki67 immunostains were photographed. Three of these were excluded from further counting as the cells were extremely crushed in 2 cases and in the other images sufficient clarity was not obtained (excluded by CP2). See Fig. 1.

Fig. 1.

Scatter plots of Ki67 percentages with a 45° line. A. Eyeballing vs. CP1. B. IR vs CP1. C. Eyeballing vs. IR. D. CP1 vs CP2.

The majority of the tumours were morphologically invasive carcinoma of no special type (89.5%). Mucinous, lobular, micropapillary, carcinoma with medullary features, apocrine, metaplastic and mixed (no special type and micropapillary) subtypes comprised the remainder. See Table 1.

Table1.

Patient Characteristics

Characteristic	Description	Number (n=201)
Race	Black	201 (100%)
Age	Range	27–86 years
	Median	52 years
	Interquartile range	42–60years
Histological type	No special type	179 (89%)
	Mucinous	7 (3.5%)
	Lobular	6 (3%)
	Micropapillary	3 (1.5%)
	Medullary features	2 (1%)
	Apocrine	2 (1%)
	Metaplastic	1 (0.5%)
	Mixed	1 (0.5%)
Histological grade	Grade 1	16 (8%)
	Grade 2	88 (44%)
	Grade 3	95 (47%)
	Not reported	2 (1%)
Intrinsic subtype	Luminal A	13 (6.5%)
	Luminal B	143 (70%)
	HER2 enriched	17 (8.5%)
	Triple negative	31 (15%)

The histological grade was assessed as grade 1 in 8%, grade 2 in 44% and grade 3 in 47%.

According to the initial biopsy reports (using a Ki67 cut-off of 14%), the intrinsic subtypes were 13 (6.5%) luminal A, 140 (70%) luminal B, 17 (8.5%) HER2 enriched and 31 (15.5%) triple negative.

Using a cut-off of 14% the number of cases classified as Luminal A by IR, CP1 and CP2 would be 9, 12 and 14 respectively. Using a Chi-square test, the p-value=0.687. (Table 2)

Table 2.

Assessment of numbers of Luminal A and Luminal B cases determined by the different methods.

		Eyeballing (n=203)	IR (n=203)	CP1 (n=202)	CP2 (n=201)	Chi-square test
14% Ki67 Cut-off	Luminal A	13 (6.4%)	9 (4.4%)	14 (6.9%)	14 (7.0%)	P=0.687
	Luminal B	143 (70.4%)	147 (72.4%)	142 (70.3%)	142 (70.2%)
20% Ki67 Cut-off	Luminal A	34 (16.8%)	22 (10.8%)	29 (14.4%)	31 (15.4%)	P=0.347
	Luminal B	122 (60.1%)	134 (66.0%)	127 (62.9%)	125 (61.7%)

Using a cut-off of 20% the number of cases classified as Luminal A by EB, IR, CP1 and CP2 would be 35, 20, 28 and 30 respectively. Using a Chi-square test, the p-value=0.347.

Comparison of eyeballing and CP1 (Fig. 2) showed a Lin’s concordance correlation coefficient of 0.698 with a 95% confidence interval of 0.629 −0.767. On a Bland Altman plot the mean difference obtained was 5.13% with 6.44% of the values outside the limits of agreement. Using a Ki67 cut-off value of 14% (Table. 3), 17 cases were classified as low (below the cut-off value for Luminal A vs. Luminal B tumours) by CP1 and 14 cases by eyeballing, and there was disagreement in 21 cases (10.4%), giving a Cohen’s Kappa statistic of ƙ=0.267 (p<0.001, fair agreement). Using a cut-off of 20%, 32 cases were classified as low by counting and 20 cases by eyeballing, and there was disagreement in 28 cases (13,9%), giving a Cohen’s Kappa statistic of k=0.387 (p<0.001, fair agreement).

Fig. 2.

Technical problems in Ki67 analysis. Ki67 immunostain at 400× magnification. A. Crush artefact precluding accurate counting. B. Pale counterstain hindering IR analysis.

Table 3.

Percentage disagreement between methods when considering all cases or only cases in the range of 10 – 25% (determined by either of the methods).

	All cases			Ki67 10–25% with either method
	Total cases	<14% cut-off	<20% cut-off	Total cases	<14% cut-off	<20% cut-off
Eyeballing vs counting	202	21 (10.4%)	28 (13.9%)	56	18 (32.1%)	25 (44.6%)
IR vs Counting	202	14 (6.9%)	21 (10.4%)	45	14 (31.1%)	21 (46.7%)
Eyeball vs IR	203	13 (6.4%)	23 (11.3%)	49	11 (22.4%)	21 (42.9%)
CP1 vs. CP2	201	8 (4.0%)	10 (5.0%)	46	8 (17.4%)	10 (21.7%)

Comparison of IR and CP1 (Fig. 2) showed a Lin’s concordance correlation coefficient of 0.790 with a 95% confidence interval of 0.744 −0.854. On a Bland Altman plot the mean difference obtained was 8.80% with 4.95% of the values outside the limits of agreement. Using a Ki67 cut-off value of 14% (Table 3), 17 cases were classified as low by counting and 9 cases by eyeballing, and there was disagreement in 14 cases (6.9%), giving a Cohen’s Kappa statistic of ƙ=0.428 (p<0.001, moderate agreement). Using a cut-off of 20%, 32 cases were classified as low by counting and 23 cases by eyeballing, and there was disagreement in 21 cases (10,4%), giving a Cohen’s Kappa statistic of k=0.560 (p<0.001, moderate agreement).

Comparison of eyeballing and IR (Fig. 2) showed a Lin’s concordance correlation coefficient of 0.716 with a 95% confidence interval of 0.650–0.783 (p<0.001). On a Bland Altman plot the mean difference obtained was −3.73 with 5.91% of the values outside the limits of agreement. Using a Ki67 cut-off value of 14% (Table 3), 14 cases were classified as low by eyeballing and 9 cases by IR, and there was disagreement in 13 cases (6,4%), giving a Cohen’s Kappa statistic of ƙ=0.403 (p<0.001, fair agreement). Using a cut-off of 20%, 20 cases were classified as low by eyeballing, and 23 cases by IR and there was disagreement in 23 cases (11,3%), giving a Cohen’s Kappa statistic of k=0.402 (p<0.001, fair agreement).

Comparison of counting done by two pathologists (Fig. 2) showed a Lin’s correlation coefficient of 0.965 with a 95% confidence interval of 0.955 – 0.974. On a Bland Altman plot the mean difference obtained was −0.76% with 4.98% of the values outside the limits of agreement. Using a Ki67 cut-off value of 14% (Table 3), 16 cases were classified as low by counting and 16 cases by eyeballing, and there was disagreement in 8 cases (4,0%), giving a Cohen’s Kappa statistic of ƙ=0.728 (p<0.001, substantial agreement). Using a cut-off of 20%, 31 cases were classified as low by CP1 and 33 cases by CP2, and there was disagreement in 10 cases (5%), giving a Cohen’s Kappa statistic of k=0.814 (p<0.001, almost perfect agreement).

When limiting comparisons to samples within the range of Ki67 readings of 10–25% with either method (Table 3), using a cut-off of 14%, eyeballing vs CP1 showed 32.1% disagreement, IR vs CP1 31.1% disagreement, eyeballing vs. IR 22.4% disagreement and CP1 vs. CP2 21.7% disagreement. Using a cut-off of 20% eyeballing vs CP1 showed 44.6% disagreement, IR vs CP1 46.7%, eyeballing vs. IR 42.9% disagreement and CP1 vs CP2 21.7% disagreement.

Discussion:

Immunohistochemistry can serve as a surrogate for molecular assessment of intrinsic subtypes in breast carcinoma, therefore many laboratories now routinely perform immunohistochemistry for oestrogen receptors, progesterone receptors, HER-2-neu and Ki67 on tissue from biopsy or excision specimens. In respect of the latter the gold standard in methodology has not yet been widely agreed. We assessed the comparability of different methods of analysis and the inter-observer agreement for the counting method performed by two pathologists.

The subtyping of the cases in this study based on immunohistochemistry, when compared to international and African/African American figures showed a lower percentage of cases classified as luminal A, and a higher percentage classified as luminal B. (7,29) Whilst this may partially reflect geographical differences, it may also be due to the cut-off level for a low Ki67 index. This underlines the fact that the most appropriate cut-off level should be determined for each laboratory specifically, (10) by correlation with molecular analysis.

The number of luminal A cases identified by each method using a Ki67 cut-off of 14% and 20% does not show a statistical difference by the Chi-square test, however, these values do not give an entirely accurate reflection of the level of agreement as they do not reflect how many of the cases identified are the same. The Cohen’s Kappa statistic is more accurate in this regard.

Overall, the best correlation was obtained between the counts done by the two pathologists (Lin’s concordance correlation coefficient 0.965). This was followed by IR vs. counting (Lin’s concordance correlation coefficient 0.790), Eyeballing vs. IR (Lin’s concordance correlation co-efficient 0.716) and then eyeballing vs. counting (Lin’s concordance correlation coefficient 0.698).

Determining a low Ki67 with a cut-off of 14%, fair agreement was obtained between eyeballing and counting, and between eyeballing and IR, with moderate agreement between IR and counting, and substantial agreement for manual counting by two observers.

Determining a low Ki67 using a cut-off of 20% yielded similar results, except that agreement for manual counting between the two observers was almost perfect.

The level of agreement was poorer when analysing those cases with Ki67 between 10% and 25%. Up to 46,7% of cases would be differently classified by IR vs. counting. Up to 44,6% of cases would be differently classified according to eyeballing compared to counting.

Even with a 4% disagreement (CP1 vs. CP2 with 14% cut-off), 8 of the 201 cases would potentially receive suboptimal management.

In practice, Ki67 determination by immunohistochemistry is often preferable to molecular methods. (20) This is especially true in resource challenged settings, where the cost of Ki67 is less prohibitive and can form part of the routine histopathological panel. An added advantage is the ability to exclude non-tumour cells such as inflammatory cells by the reporting pathologist.

Whilst the recommended method of determining the Ki67 index is manual counting, there is no definite “gold standard”.

Eyeballing is the most subjective method of analysing Ki67, and the values are usually rounded to the nearest 10, which reduces the accuracy. However, this is the most convenient method to use and some authors claim that it is more accurate than other methods because it is based on an overall impression of all the tissue stained. (28)

Zabaglo et al. identified the elimination of subjectivity in the interpretation of the Ki-67 index as an advantage of image analysis. (30) The major problem encountered in our study with the use of IR was the intensity of the counterstain. The intensity of the counterstain was not adequate in some cases and could not be corrected by altering the degree of contrast especially in the older cases examined. Many of the negative cells were not detectable by the software. If this method were to be used routinely, the staining characteristics would have to be optimised for this purpose. Tumour cells are not always readily distinguishable from non-neoplastic cells such as endothelial, stromal and inflammatory cells using IR, although the size of the tumour cells is used as a guideline.

Counting is tedious and time consuming, and distractions are problematic. Intra-tumoural heterogeneity can also affect the results obtained, although it is less of a problem in core biopsies than in tumour blocks as the volume of tumour is less. This factor was eliminated in this study by the use of identical photomicrographs that resulted in the same fields being counted by both observers. Care must be taken to count only invasive cells, avoiding areas of in situ neoplasia. The cells to accept as positively staining may also differ between observers, such as very pale staining cells, speckled staining and cells that are partially overlapping. It has been recommended that any staining should be accepted as positive, since differences from one cell to another reflect variable staining patterns in different parts of the cell cycle. (31) Varga et al. found that despite attempting to better define which nuclei should be accepted as positive, there was no improvement in interobserver agreement. (28)

Limitations

Subjectivity in selection of hot spots for slide imaging is a limitation since the areas are selected by individual pathologists.

The use of core biopsies in this study is a potential limiting factor, but counting of 500 cells was an aspect of the methodology in this study, and where this was not possible the core was excluded. Immunohistochemistry is carried out on core biopsies prior to administration of neo-adjuvant therapy, and the resection specimens often have limited residual tumour post-chemotherapy.

Conclusion

The different methods of assessing Ki67 yield acceptable correlation between results at high levels and low levels. The intermediate levels (10–25%) are problematic and using different methods results in discrepancies that are particularly concerning in the subtyping of luminal cancers into Luminal A and Luminal B subtypes. Ideally, these cases should be further analysed by molecular methods to assist with precise classification, to determine the Ki67 cut off value per laboratory and to assess which Ki67 determination method is most superior. Further development in molecular methods may result in these becoming more practical in the future.

Abbreviations

IR

Immunoratio©

CP1

Counting by pathologist 1

CP2

Counting by pathologist 2

P1

Pathologist 1

P2

Pathologist 2

EB

Eyeballing

DIA

digital image analysis