This study included 512 consecutive patients with primary invasive breast cancer treated and diagnosed at Malmö University Hospital between 1^st January 1988 and 31^st December 1992 [16-18]. The median age at diagnosis was 65 (range 27-96) years and median follow-up time to first breast cancer event was 128 months (0-207). Information regarding the date of death was obtained from the regional cause-of-death registries for all patients. Complete treatment data were available for 379 (76%) patients, 160 of whom had received adjuvant tamoxifen. Twenty-three patients received adjuvant chemotherapy. Two hundred patients received no adjuvant systemic treatment. Ethical permission was obtained from the Local Ethics Committee at Lund University (Dnr 613/02), whereby informed consent was deemed not to be required other than by the opt-out method.

TMAs were constructed using two 0.6 mm cores taken from areas representative of invasive cancer and mounted in a recipient block using a manual arraying device (MTA-1, Beecher Inc, WI, USA) as previously described [19]. TMA blocks were stored in Dept. of Pathology in Malmö University Hospital, Sweden. TMA sections were cut immediately prior to staining.

Sections (4 μm) were dried, deparaffinised and rehydrated through descending concentrations of ethanol. Heat-mediated antigen retrieval was performed using microwave treatment for 2 × 5 min in a citrate buffer (pH 6.0), before being processed either in the Ventana Benchmark system (Ventana Medical Systems Inc, AZ) using pre-diluted antibodies to ER (Anti-ER, clone 6F11), progesterone receptor (PR) (Anti-PgR, clone 16) and Her2 (Pathway CB-USA, 760-2694) or in the Dako Techmate 500 system (Dako, Glostrup, Denmark) for Ki-67 (1:200, M7240; Dako) and survivin (1:50, D-8 Santa Cruz, CA). ER, PR and Ki-67 were quantified using a commercially available automated nuclear algorithm (IHC-MARK; OncoMark Limited, Ireland), as previously described [20]. ER, PR and Ki-67 negativity was defined as < 10% positively stained nuclei. Her2 staining was evaluated according to a standard protocol (HercepTest) and scored as 4 intensities (i.e. negative, weak, moderate and strong), namely 0-3+; these scores were divided into two groups, with negative to weak (0-2+) Her2 expression (Her2^-) and strong (3+) overexpression (Her2⁺).

The Aperio ScanScope XT Slide Scanner (Aperio Technologies, Vista, CA) system was used to capture whole slide digital images with a 20× objective. Digital images were managed using Spectrum (Aperio). Tumour and stromal elements were identified using Genie (Aperio) pattern recognition software and a quantitative scoring model for both nuclear and cytoplasmic survivin was developed using the positive pixel count algorithm (Aperio).

Spearman's Rho correlation was used to estimate the relationship between duplicate cores from individual tumours. Differences in distribution of clinical data and tumour characteristics between samples with a high and low survivin nuclear autoscore (SNAS) and CNR were evaluated using the χ² test. Kaplan-Meier analysis and the log rank test were used to illustrate differences between OS and BCSS according to survivin expression. Cox regression proportional hazards models were used to estimate the relationship to OS and BCSS of survivin, patient age, lymph node status, tumour grade, Her2, PR, and ER status in the patient cohort. For decision tree analysis, all patients were randomly divided into 10 subsets. A decision tree model was selected using a 10-fold cross-validation approach. Ten consecutive decision tree models were independently constructed using the SNAS and CNR continuous output from 9 subsets. Prognostic accuracy of each decision tree model was tested using the remaining set of patients, with the model displaying the highest accuracy being selected as the optimal model for the dataset. All calculations were performed using SPSS version 12.0 (SPSS Inc, Chicago, IL). A p-value < 0.05 was considered statistically significant.

The specificity of the anti-survivin antibody was confirmed in our previous study [15]. Following optimisation of the staining procedure, it was possible to evaluate survivin protein expression in 70.1% (n = 359) of the tumours represented on the TMA. To evaluate our study for any potential selection bias, baseline clinicopathological characteristics from both the evaluated ("survivin known") cohort (n = 359) and the unevaluated ("survivin unknown") cohort (n = 153) are presented in Table 1. This illustrates that the clinicopathological characteristics were well matched in both evaluated and unevaluated cohorts.

In order to develop an accurate quantitative automated model of survivin expression, a pattern recognition algorithm developed in Genie (Aperio) was used to identify tumour from stroma and slide background. This algorithm was trained using five prospectively selected patterns - 1) tumour positive, 2) tumour negative, 3) stromal fibroblast and extracellular matrix, 4) lymphocytes and 5) slide background (Figure 1A). Following manual annotation of the training set, the five patterns were provided as an input for the pattern recognition algorithm in order to select the optimum approach. Performance of the pattern recognition algorithm was then tested in 20 randomly selected tissue cores (validation set). Comparison of automated quantification to manual annotation at pixel level resolution within the validation set revealed that the automated approach had an accuracy of 87.3%. Additional file 1 illustrates the training and output from the pattern recognition algorithm.

The positive pixel count (Aperio) was then applied to develop a quantitative scoring model for tumour-specific nuclear and cytoplasmic survivin expression. A pseudo-colour "mark-up" image was generated for each core as the algorithm output (Figure 1B). The accuracy of the algorithm was examined in high power fields (Figure 1B) and was deemed acceptable by a histopathologist.

The algorithm was used to calculate total survivin intensity, as well as cytoplasmic and nuclear intensity for each core. One of the issues to consider when examining staining intensity alone is that even a small fraction of staining artifact can significantly alter intensity values. Therefore, to reduce noise secondary to staining artifact SNAS (which combines the product of percentage positive nuclei and nuclear intensity) was proposed as an alternative automated scoring model for evaluating nuclear survivin expression. A combined percentage and intensity autoscore was calculated for both nuclear and cytoplasmic survivin expression, as previously described [15]. A strong correlation was evident between duplicate cores from individual tumours for nuclear percentage (Spearman's Rho = 0.350; p < 0.001), nuclear intensity (Spearman's Rho = 0.479; p < 0.001), cytoplasmic percentage (Spearman's Rho = 0.620; p < 0.001) and cytoplasmic intensity (Spearman's Rho = 0.619; p < 0.001), indicating a homogeneous pattern of expression. As tumours were arrayed in duplicate, the maximum value for each tumour was used for further analysis.

We previously described a relationship between the CNR of survivin and patient outcome [15]; however, the relationship between nuclear and cytoplasmic survivin expression within individual tumours was not investigated. Comparison of nuclear and cytoplasmic autoscores from individual tumours revealed a good correlation between maximum nuclear and cytoplasmic autoscore (Spearman's Rho = 0.753; p < 0.001), suggesting that nuclear and cytoplasmic expression co-exist. In an attempt to further evaluate the relationship between nuclear and cytoplasmic survivin expression, patients were divided into groups based on SNAS centiles (Figure 1C). This revealed a close relationship between nuclear and cytoplasmic survivin. When nuclear survivin was low, cytoplasmic survivin was also low; likewise, when nuclear survivin was high, cytoplasmic survivin was also high. However, a divergence between the two measures in the mid range values suggests that a ratio such as the CNR may be useful.

Decision tree analysis was then used to examine the prognostic role of nuclear and cytoplamic survivin expression. Using a 10-fold cross-validation approach, the normalised importance of SNAS, survivin cytoplasmic autoscore and CNR was computed in relation to BCSS and OS (Figure 1D). SNAS was consistently the most prognostic marker for OS and BCSS followed by CNR, again suggesting an important prognostic role for nuclear survivin.

In our initial study, we reported no link between survivin nuclear or cytoplasmic intensity and survival; however, an increased SNAS was associated with a reduced BCSS [15]. Application of the same SNAS threshold (8) in this cohort demonstrated no association between SNAS and either BCSS (p = 0.123) or OS (p = 0.298). However, univariate cox regression analysis of SNAS as a continuous variable revealed an association between SNAS and a decreased BCSS (HR 1.028, 95% CI 1.001-1.056; p = 0.044). A decision tree model was thus employed to identify the optimal threshold for SNAS in this cohort (Figure 2A) and predicted 4.26 as an optimal cut-off for automated SNAS (Additional file 2).

Kaplan-Meier analysis comparing low (n = 158) and high (n = 201) SNAS, as defined by decision tree analysis, revealed that a high SNAS was associated with decreased OS (p = 0.003) and BCSS (p = 0.004) (Figure 2B). Cox univariate analysis demonstrated that SNAS was a predictor of a reduced BCSS (HR 2.05, 95% CI 1.27-3.31, p < 0.001) and OS (HR 1.51, 95% CI 1.13-2.01, p = 0.005). Multivariate Cox regression analysis controlling nodal status, grade, ER status, PR status, Her2 status, tumour size and age revealed that SNAS was an independent predictor of OS (HR 1.68, 95% CI 1.22-2.32, p = 0.002) and BCSS (HR 1.84, 95% CI 1.10-3.08, p = 0.02), along with lymph node status, grade and age (Table 2). The relationship between SNAS and other clinicopathological variables was also examined (Table 1). A high SNAS was associated with high grade (p = 0.005) and Ki-67 positivity (p = 0.010). No relationship was evident between SNAS and patient age, nodal status, hormone receptor expression or Her2 status.

Having demonstrated a significant relationship between SNAS and survival, the CNR of survivin protein expression was examined. The initial study used a non-supervised approach to identify a CNR of 5 to divide patients into groups with low and high CNR [15]. Decision tree models were again used to identify an optimal threshold for CNR. This model predicted a cut-off of 4.87 (Figure 2C) as the optimal threshold for CNR, which correlated well with the value used in the previous study [15]. There was an excellent correlation between low CNR and high CNR defined using a threshold of 4.87 or 5 (R = 0.983, p < 0.001). Patients were thus divided into groups of low and high CNR using 5 as a threshold on the automated continuous scores (Additional file 2).

Forty-nine percent (n = 161) of tumours had a CNR greater than 5. Kaplan-Meier analysis of OS and BCSS revealed that a survivin CNR of < = 5 was associated with a reduced BCSS (p = 0.005) but had no effect on OS (p = 0.187) (Figure 2D). Univariate Cox regression analysis (Table 2) confirmed the relationship between CNR and an improved BCSS (HR 0.49, 95% CI 0.29-0.81, p = 0.006). Multivariate analysis controlling for age, tumour size, nodal status, grade, ER, PR and Her2 status revealed that an increased survivin CNR was a predictor of a prolonged BCSS (HR 0.47, 95% CI 0.27-0.82, p = 0.008), along with tumour grade and lymph node status (Table 2). The relationship between survivin CNR and other clinicopathological variables was also examined. An increased survivin CNR was associated with ER-positive (p = 0.045), low grade (p = 0.007), Ki-67-negative (p = 0.001) and Her2-negative (p = 0.026) tumours (Table 1).

Given the previously described relationship between increased CNR and ER positivity, we proceeded to examine the relationship between nuclear survivin and outcome in ER-positive patients. Nuclear survivin was measured as an increased SNAS or a low CNR. A low CNR (<5) was associated with a decreased BCSS (p = 0.029) in ER-positive patients (n = 273) (Figure 3A), an effect that was not evident in ER-negative (n = 81) patients (p = 0.062). A low CNR (<5) was also associated with a trend towards a decreased OS (p = 0.062) in ER-positive patients (Figure 3A). A high SNAS was also associated with a reduced BCSS (p = 0.005) and OS (p = 0.013) (Figure 3B) in ER-positive patients and a reduced BCSS (p = 0.036), but not OS (p = 0.136) in ER-negative patients.

Having demonstrated that nuclear survivin predicts outcome in ER-positive patients, we proceeded to examine its effect on ER-positive patients who received tamoxifen (n = 125). This revealed that a high SNAS was associated with a reduced OS (p = 0.006) and BCSS (p = 0.013) (Figure 4A) in tamoxifen-treated ER-positive patients. Likewise, a low CNR was associated with a reduced OS (p = 0.021) and BCSS (p = 0.018) in ER-positive patients who received tamoxifen (Figure 4B). Multivariate Cox regression analysis demonstrated that neither SNAS (HR 1.87, 95% CI 0.71-4.88, p = 0.194) nor CNR (HR 0.41, 95% CI 0.15-1.13, p = 0.087) were independent predictors of BCSS in tamoxifen-treated patients; however, a low CNR was an independent predictor of OS in tamoxifen-treated patients (HR 0.44, 95% CI 0.23-0.87, p = 0.018), after controlling for age, tumour size, nodal status, grade, PR and Her2 status (Table 3).