Novel image analysis approach for quantifying expression of nuclear proteins assessed by immunohistochemistry: application to measurement of oestrogen and progesterone receptor levels in breast cancer

The oestrogen receptor (ER) remains the only reliable predictor of endocrine responsiveness in breast cancer, and is arguably the single most important predictive biomarker in clinical oncology today [1]. Moreover, one of the most studied ER-regulated genes is the progesterone receptor (PR). Approximately 70% to 80% of all invasive breast cancers are ER-positive and thus are considered likely to respond to endocrine therapy. The PR, which is positive in approximately 60% of cases, may be even more important in predicting response to anti-oestrogens [2].

Both premenopausal and postmenopausal women benefit from 5 years of treatment with the anti-oestrogen tamoxifen [3]. Current treatment guidelines for premenopausal women with hormone-responsive breast cancer advocate a combination of ovarian ablation/suppression or chemotherapy, followed by 5 years of tamoxifen treatment [4‐6]. In hormone-responsive postmenopausal women, data from large prospective randomised controlled trials involving aromatase inhibitors are now emerging and herald new standards in adjuvant endocrine treatment [7, 8]. The International Expert Consensus on the Primary Therapy of Early Breast Cancer states that tamoxifen may be an acceptable option, but that aromatase inhibitors have shown superiority over tamoxifen in postmenopausal breast cancer [5].

Irrespective of their menopausal status, arguably the single most important issue for any breast cancer patient is the assessment of her tumour hormone receptor status. The hormone receptor status is routinely evaluated in all resected primary tumours to assess the levels of ER and PR. Inmmunohistochemistry (IHC) performed on formalin-fixed tissue sections is now the most commonly used assay, having replaced biochemical-based methods. The hormone receptor status is currently assessed by a pathologist; with a cutoff threshold of 10% positive tumour cells being commonly employed to predict responsiveness to adjuvant hormonal therapy. Such a threshold can lead to significant intra-observer variability. For example, one study of 172 German pathologists highlighted the difficulties that can arise from manual assessment, with 24% of ER staining interpreted as being falsely negative [9]. Improved image analysis technologies have the potential to circumvent the burden of interpretation and intra-observer variability, offering the potential to develop objective automated quantitative scoring models for IHC. A move away from the semiquantitative manual scoring models currently employed should lead to less variability in results, to increased throughput and to the identification of new prognostic subgroups, which may not have been evident following initial manual analysis alone [10].

In the present article we propose an automated approach, based on unsupervised learning, to accurately assess the ER and PR expression levels in an extensive cohort of breast cancer specimens. In particular, our approach employs a novel approach to the identification of tumour nuclei, whereby nontumour structures, including stromal components and lymphocytic infiltrate, are automatically excluded from any analysis. Such an approach should allow for more accurate assessment of the IHC signal.

The present study involved the development of an automated unsupervised nuclear algorithm to objectively assess the ER and PR expression levels in 743 breast tumours. Two breast cancer cohorts (n = 743) as described above were used (Table 1). Manual ER and PR expression data were available for 654 (88%) patients and 640 (86%) patients, respectively. As a number of cores were not suitable for image analysis due to large image artefacts, automated analysis was restricted to 639 patients (86%) for whom ER data were available and to 622 (84%) patients for whom PR data were available.

The purpose of the present study was to develop an unsupervised algorithm for automated quantification of IHC-determined nuclear protein expression in breast tumour specimens. In the current article we present data on the automated quantification of the ER and the PR in 743 primary breast tumours using such an algorithm. Our approach differs from other commercially available packages in that it does not require prior data for training purposes, and it uses a novel, fully automated approach to isolate tumour cells from stromal tissue and lymphocytic infiltrate. A good counterstain is helpful; however, it does not have to be specifically controlled for automated analysis. Heterogeneity in respect to counterstaining is handled in our approach by using a particular colour space (CIE Luminance U-chromatic component 1 V-chromatic component 2) combined with different morphological image operations. Such an approach allows for the identification of negative tumour nuclei from stromal elements, thus allowing for a more precise determination of protein expression. The accuracy of this approach was evident from the excellent correlation seen between automated quantification and manual analysis (Figure 1).

The first attempts to use automated methods to quantify protein expression, as assessed by IHC, were undertaken almost two decades ago [14, 15]. Until recently, however, the image quality and analysis software did not permit high-throughput automated assessment of histological slides. A number of groups have now published data on the automated assessment of the ER and the PR in breast cancer [16‐18]. In general, these studies are in agreement with the data presented in the current study in relation to describing an excellent correlation between manual and automated analysis. The studies by Turbin and colleagues and by Gokhale and colleagues describe commercially available algorithms that require specific input and training by a pathologist [16, 17]. The algorithm described in the current study differs from these approaches in that it is entirely unsupervised and does not require any a priori data. Other approaches to the quantification of the ER and the PR described in the literature have used antibody-conjugated fluorophores and fluorescent microscopy systems [19, 20]. This method, however, requires users to perform specific and complex staining protocols, which may not be suitable for a busy diagnostic service.

There has been much discussion in the literature regarding the optimal thresholds for the ER and/or the PR when identifying patients who will benefit from anti-endocrine therapy [20‐23]. Currently, a threshold of 10% positive cells is commonly used to predict a patient's response to adjuvant hormonal therapy. Such a cutoff threshold can lead to significant intra-observer variability. One of the aims of the present study was to attempt to validate currently used thresholds for the ER and the PR, as well as to find new thresholds, using an unsupervised clustering approach. To this end, we were able to utilise a cohort of patients who had participated in a prospective randomised control trial comparing tamoxifen treatment with no adjuvant treatment [2]. The untreated arm of the trial was used to evaluate the prognostic value of new cutoff thresholds, while the treated arm was used to evaluate the predictive power of such thresholds.

RFC, an approach that has been used previously by our group [10], was utilised to identify new thresholds for survival analysis. This approach identified optimal thresholds of 7% for the ER and 5% for the PR. Interestingly, the 7% threshold for the ER was associated with a similar outcome in Cohort I and an improved outcome when compared with a manual 10% cutoff level in Cohort II (Table 2). This effect was not evident for the PR.

The true prognostic power of a 7% threshold was investigated in the untreated arm of Cohort II, and this confirmed our findings regarding both markers – specifically, ER manual analysis using a 10% cutoff threshold was associated with a HR of 0.64 (95% confidence interval = 0.46 to 0.81, P < 0.001). A 7% threshold as determined following RFC, however, was associated with a lower HR and narrower confidence intervals (HR = 0.59, 95% confidence interval = 0.44 to 0.77, P < 0.001), Finally, we examined the ability of our clusters to predict the tamoxifen response in a premenopausal cohort from a randomised control trial. This demonstrated that the RFC-based approach was associated with better outcomes, indicating that this approach would be a more accurate predictor of tamoxifen response (Table 3). This observation suggests that 10% is a suboptimal threshold for the ER when predicting the tamoxifen response. What should also be acknowledged, however, is that a 7% cutoff threshold for manual analysis is perhaps not feasible and would lead to an increase in both intra-observer and inter-observer variability.

These data also confirm previously published findings describing the heterogeneous pattern of expression for the PR [20]. One of the arguments against TMA-based approaches is that they do not give a true representation of proteins that display such heterogeneity in expression terms [24, 25]. In the present study we have used an image analysis approach to address marker heterogeneity. If TMAs are to be used for high-throughput biomarker validation in the future, an ability to identify proteins exhibiting such varied expression patterns would be extremely helpful and avoid the reporting of potentially biased data. The method would also allow investigators to proceed, in a systematic manner, onto full-face sections at an earlier stage in the validation process. Our approach revealed that the ER displayed a much more homogeneous pattern of expression than the PR (Figure 1g, h). Our data are in agreement with other studies. For example, in a review of 5,993 cases, Nadji and colleagues demonstrated that the ER displayed a homogeneous pattern of expression in 92% of cases, whilst the PR displayed a heterogeneous pattern of expression in 21% of cases [26]. We believe this observation may indicate why our results for the PR were not as consistent as those seen for the ER; however, this does not preclude TMA-based analysis of PR expression.

We have successfully developed and validated a novel approach to quantify nuclear protein expression in breast tumours using image analysis of IHC data. Moreover, we have validated 7% as the optimal cutoff threshold for the ER, by utilising two independent cohorts comprising a total of over 700 patients. Our data would also suggest that image analysis approaches may provide more sensitive measures of predicting tamoxifen response, whereby a 7% cutoff threshold predicts improved response to tamoxifen. Based on our findings in the present study, 13 patients (3%) who would not currently receive tamoxifen would have benefited from the drug. One potential weakness of the current study is that it was based on a TMA platform. Therefore, it is planned to proceed to apply this technology to full-face sections derived from breast biopsies and resections. This would allow the algorithm to be utilised in routine pathological practice.

There is an urgent need for new prognostic and predictive assays that would allow for improved patient stratification in breast cancer. Whilst a huge emphasis has been placed on DNA microarrays as the basis of such new tests [27], it is now becoming obvious that IHC surrogates may be a more appropriate clinical assay than gene expression-based platforms in the future [28]. As demonstrated by Nielsen and colleagues and by Careyand colleagues, it is possible to identify molecular subgroups (luminal A, luminal B, basal and Her2) using a small number of IHC markers [29, 30]. Likewise, Ring and colleagues have reported on a novel panel of five antibodies that can predict outcome in ER-positive tumours [31]. More recently, Crabb and colleagues described a novel eight-marker, IHC-based, prognostic test for patients with advanced lymph-node-positive (>4 positive lymph nodes) disease [32]. These findings combined with large-scale antibody-based-proteomics resources, such as the Human Protein Atlas programme [33], will most probably lead to an increased use of multiplex IHC assays as both predictive and prognostic tests. The use of automated algorithms such as the one described in the present paper can only help to advance this process, whilst the use of novel approaches such as RFC may lead to the identification of new prognostic and predictive subgroups, allowing for tailored treatment regimens for individual patients – also known as personalised medicine.