Prognostic value of automated KI67 scoring in breast cancer: a centralised evaluation of 8088 patients from 10 study groups

This analysis was conducted within the Breast Cancer Association Consortium [25], which is a large, ongoing collaborative project involving study groups across the globe. Figure 1 shows that we collected a total of 166 TMAs containing 19,039 cores, representing 10,005 patients from 13 study groups (Additional file 1: Table S1). Ten study groups provided unstained TMAs which were then stained and digitised in the Breakthrough Core Pathology laboratory at the Institute of Cancer Research (ICR) and the academic biochemistry laboratory of the Royal Marsden Hospital (RMH), London, UK. Two groups (MARIE and PBCS) provided pre-stained TMAs which were also digitised in our centre. One study (SEARCH) provided TMA images acquired using a similar Ariol technology (a digital image acquisition and analysis system) to the one adopted for this analysis. Of the 10,005 patients, 1917 were excluded on account of failing predefined quality control checks (N = 946) or due to absent data on follow-up times and/or vital status (N = 971). As a result, a total of 8088 patients from 10 study groups with a median follow-up of 7.5 years and a total of 1401 breast cancer specific deaths were used in the survival analysis involving automated KI67. Of these, 2440 patients with pathologists’ visual KI67 scores in addition to automated KI67 scores were used to extrapolate a visual from an automated cut-off point, following which comparative survival analyses involving visual and automated KI67 scores were conducted. Information on other clinico-pathological characteristics of tumours including histological grade, nodal status, tumour size, stage, adjuvant systemic therapy (endocrine therapy and/or chemotherapy) and other IHC markers (i.e. oestrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER2)) were obtained from clinical records. Additional Ariol HER2 data were obtained for a subset of patients with missing clinical HER2 data but for whom data on ER and PR were available (N = 403). All patients provided written informed consent and all participating studies gained approval from the local ethical committees and institutional review boards.

Sections were dewaxed using xylene and rehydrated through graded alcohol (100, 90 and 70 %) to water. Slides were then placed in a preheated (5 min, 800 W microwave) solution of Dako Target Retrieval solution pH 6.0 (S1699), microwaved on high power for 10 min and then allowed to cool in this solution at room temperature for 10 min. In the next stage, the slides were placed on a Dako autostainer and stained using a standard protocol with Dako MIB-1 diluted 1/50 and visualised using the Dako REAL kit (K5001). The MIB-1 antibody was also adopted for the staining of the TMAs that were not part of those centrally stained at the ICR, but at varying concentrations (PBCS = 1:500; MARIE = 1:400 and SEARCH = 1:200) (Additional file 1: Table S2).

KI67 scoring has been described previously [26], but briefly all TMAs were digitised using the Ariol 50s digital scanner. Fifteen TMAs were selected as a training set. These were scored visually by a pathologist (MA) using a computer-assisted visual (CAV) counting method and used to validate the automated method. The CAV method relied upon built-in features of the Ariol digital system to count negative and positive nuclear populations within 250 μm × 250 μm squares separated by grids. The standard CAV approach entailed the counting of at least 1000 cells across the entire spectrum of each core. In the majority of cores, more than 1000 cells were counted even though fewer than this number was counted in a small minority. Overall, cores with more than 500 cells were considered to be of satisfactory quality. The CAV method is precise, prevents double counting and was observed to have excellent intra-observer reproducibility when a random subset of cores (N = 111) were re-scored at an interval of 3 months from the first time they were scored (observed agreement/kappa = 96 %/0.90); good core-level agreements with two other independent scorers (observed agreement/kappa: CAV vs scorer 2 = 87 %/0.66; CAV vs scorer 3 = 84 %/0.59; scorer 2 vs scorer 3 = 89 %/0.69) were also observed in a randomly selected subset of 202 cores. Visual scoring in the external TMAs involved both quantitative and semi-quantitative methods. Each core from each patient was scored by two independent pathologists and the KI67 score for each patient was then taken as the average score from the two scorers across all cores for that patient.

The automated scoring was performed using the Ariol machine which has functionality that allows for the automatic detection of malignant and non-malignant nuclei using shape and size characteristics. Using colour deconvolution, it also distinguishes between DAB-positive and DAB-negative (haematoxylin) malignant cells. To determine the negative and positive populations of cells, an appropriate region of interest of the malignant cell population in a core was demarcated and two colours were selected to indicate positive and negative nuclear populations. The appropriate colour pixels were then selected to represent the full range of hue, saturation and intensity that was considered representative of the positive and negative nuclear classes [26]. Subsequently, the best shape parameters that discriminated malignant and non-malignant cells according to their spot width, width, roundness, compactness and axis ratio were also selected. The data were divided into a training and a validation subset and the automated and visual scoring for KI67 showed good agreement (observed agreement = 87 %; Kappa = 0.64) and discriminatory accuracy (AUC = 85 %) in the validation subset, hence allowing for the adoption of this method for the scoring of all 166 TMAs.

For patients with multiple cores from the same tumour, we used the average KI67 score across valid cores to represent the % positive cells in that tumour. Descriptive analyses of the distribution of KI67 according to clinical and pathological characteristics of the patients were conducted using the non-parametric Kruskal–Wallis equality of medians test for continuous measures and the paired chi-squared test for categorical measures. The relative survival probabilities for patients in different quartiles of the KI67 distribution were compared using Kaplan–Meier survival curves for the 10-year breast cancer specific survival (BCSS). To allow for prevalent cancers, time at risk was left-censored for study entry. It was decided, a priori, not to make any assumptions on a prognostic cut-off point for automated KI67 scores in our dataset but instead to leverage on the continuous values to observe a prognostic threshold. As a result, we performed quartile analysis by dividing the continuous KI67 scores into quartiles (Q1–Q4) and examining the prognostic differences among the different quartiles for all patients in the study. The 10-year BCSS was determined using Kaplan–Meier survival curves and Cox-proportional hazards regression models stratified by ER status (positive vs negative) and according to nodal status (positive vs negative) and other IHC markers. The univariate Cox models were partially adjusted for study group and age at diagnosis while the multivariate models had further adjustments for other known prognostic factors including histological grade, tumour size, nodal status, morphology, ER, PR, HER2 and adjuvant systemic therapy (endocrine and/or chemotherapy). In the multivariate models, missing values for other covariates were addressed using the multiple imputation plus outcome (MI+) approach [27]. Because of observed violation of the proportionality assumption of the Cox model by automated KI67, it was modelled as a time-varying covariate using an extension of the Cox model that allows for the inclusion of a coefficient (T) that varied as an exponential function of time. The log of the coefficient is indicative of both the direction and the magnitude of change in hazard ratio with time, such that if log T < 1 then hazard falls with time, while if log T > 1 then hazard increases with time. Known violation of the proportional hazards assumption by ER was addressed in the same way. Consistency of hazard ratio (HR) estimates across the different study groups was evaluated using the I ² statistic, derived by performing a fixed-effect meta-analysis of study-specific HR estimates. To enable direct comparison between the visual and automated KI67 scores, we extrapolated a visual from an automated cut-off point in a linear regression model and used the resulting cut-off point for all further analyses. All analyses were conducted using STATA statistical software version 10 (StataCorp, College Station, TX, USA). Statistical tests were two-sided and P < 0.05 was considered statistically significant.

In all, a total of 143 TMAs containing 15,313 cores from 8088 patients were used in this analysis, as shown in Fig. 1. The studies included in this analysis used different TMA designs (Table 1). More than half (4431/55 %) of the patients had KI67 scores on at least two cores and evaluation of dichotomous categories revealed concordant KI67 status in 83.7 % of the patients. When we examined the distribution of continuous KI67 scores among categories of the different clinical and pathological characteristics we observed this to differ according to histological grade, tumour size, morphology, ER status, PR status and HER2 status, but not nodal status or stage at diagnosis (Fig. 2). The distribution of these characteristics for patients with high KI67 (Q4 or >12 % positive cells) and low KI67 (Q1–Q3) are shown in Additional file 1: Tables S3 and S4 for ER-positive and ER-negative patients, respectively.

Using continuous measures of KI67 categorised into quartiles, we observed poorest survival in the highest quartile, corresponding to 12 % positive cells, but little difference in survival between the other three (Q1–Q3) quartiles (log-rank P = 1.2 × 10⁻⁵; Fig. 3a). As a result, the continuous KI67 value was dichotomised at the threshold of 12 % in subsequent analyses. High KI67 was significantly associated with worse 10-year BCSS overall (log-rank P = 3.1 × 10⁻⁷) among ER-positive cancers (log-rank P = 1.3 × 10⁻³) but not ER-negative cancers (log-rank P = 0.35) (Fig. 3b–d, respectively). Similarly, in multivariate models, high KI67 expression was significantly associated with worse 10-year BCSS among ER-positive cancers (HR at baseline = 1.96; 95 % CI = 1.31–2.93) but not ER-negative breast cancers (HR = 1.23; 95 % CI = 0.86–1.77; P-heterogeneity = 0.064) (Table 2). Further stratification of ER-positive cancers according to nodal status showed that high KI67 was associated with worse survival in both node-negative and node-positive cancers in multivariate analysis (node-negative 2.47 (1.16–5.27); node-positive 1.74 (1.05–2.86); P-heterogeneity = 0.67) (Table 2). The association between KI67 and survival was significant among ER-positive patients who did not receive chemotherapy (1.95 (1.18–3.21); P = 0.009) but not among those who did (1.89 (0.84–4.29); P = 0.124; P-heterogeneity = 0.60). We found no evidence of between-study heterogeneity in estimates of HR for ER-positive patients (I ² = 0.0 %, P = 0.94) or ER-negative patients (I ² = 0.0 %, P = 0.86) (Additional file 2: Figure S1). Among hormone receptor-positive breast cancers, the HR for KI67 was not significantly different according to HER2 status (Table 2; P-heterogeneity = 0.270). Modest evidence for a poorer prognosis among high, relative to low, KI67 was also seen for triple-negative breast cancers (1.70 (1.02–2.84); P = 0.04). No significant associations with prognosis were found for KI67 among HER2-positive (i.e. ER^–/PR^–/HER2⁺) breast cancers (0.91 (0.60–1.36)) (Table 2).

The automated cut-off point of 12 % positive cells corresponded to a visual cut-off point of 24.2 % based on a linear regression model comprising patients with quantitative data on both methods. The visual cut-off was rounded up to a cut-off point of 25 %. Strong evidence (P < 0.0001) in support of a positive linear correlation (r = 0.63) between automated and visual scores was observed and continuous automated scores showed good discriminatory accuracy against the visually determined binary classes (AUC = 82 %, 95 % CI = 80–84 %)(Additional file 3: Figure S2). Twenty-six percent of the patients were classified as having high visual KI67, in contrast to 29 % for the automated KI67 scores; cross-classification of visual and automated categories revealed better specificity (84 %) than sensitivity (65.6 %) for the automated score in classifying visually determined categories (Additional file 1: Table S5). High KI67 was associated with worse survival in Kaplan–Meier curves based on both automated (log-rank P = 9.8 × 10⁻⁶) and visual (log-rank P = 3.8 × 10⁻¹⁴) KI67 scores even though attenuation of the difference between strata was observed for automated KI67 scores (Additional file 4: Figure S3). In two separate models for visual and automated KI67 scores each adjusted for age at diagnosis and study group we observed stronger evidence for an association between KI67 and survival for the visual KI67 score than for the automated KI67 score (Table 3). Analysis of model fit revealed similar parameters for both scores, however, especially in ER-positive breast cancers (AIC/BIC: visual = 2656/2618; automated = 2675/2638) (Table 3). When we performed further adjustments for other prognostic factors in multivariate Cox models of imputed datasets, we observed both visual and automated KI67 scores to be significantly associated with survival for all patients (HR (95 % CI): visual = 1.75 (1.23–2.49); automated = 1.61 (1.14–2.28)) and for ER-positive patients (visual = 2.30 (1.34–3.94); automated = 2.10 (1.28–3.47)), but not for ER-negative patients (visual = 1.63 (0.97–2.72); automated = 1.28 (0.79–2.05)) (Table 3).