Development and evaluation of a virtual microscopy application for automated assessment of Ki-67 expression in breast cancer

With the emergence of virtual microscopy and whole slide scanning techniques, there is an increasing need for efficient tools to automate assessment of digitized biological samples. One possible solution is to integrate computer vision methods with a virtual microscopy platform and to run the image analysis software on the same server system as the virtual slides are stored.

A considerable number of published scientific studies have addressed computer vision for quantification of protein expression as determined by immunohistochemistry (IHC) [1‐16]. Only one of the previous studies is based on an open source solution [17]. Very few studies have compared human visual interpretation and computer vision of IHC expression levels with regard to clinically important endpoints, such as disease outcome [2, 15, 16].

While tissue sample processing and IHC staining methods are increasingly automated, the evaluation of staining results is still predominantly performed by visual assessment. A human interpreter has excellent image comprehension and pattern recognition capabilities, but is prone to substantial variability in quantification tasks. Computer vision methods are capable of processing images consistently and generally perform well in repetitive processes. Virtual microscopy combined with computer vision techniques can aid the human observer by analysis of large tissue areas at a high magnification. The digital sample (i.e. the virtual slide) can be an entire section of a single cancerous tumour or an array of 100-200 tumour tissue samples assembled by the use of tissue microarray technology [18].

We decided to develop and study a computer vision method for IHC analysis that can be run on a virtual microscopy platform and to compare the method to visual interpretation of IHC staining. A highly studied biomarker, Ki-67, with known prognostic value in many cancer forms was chosen as the target [9, 11, 12, 19‐22]. Ki-67 is a protein associated with cell proliferation and is present in all other cell cycle phases except G0, the resting phase. Ki-67 is thoroughly studied in breast cancer and Ki-67 immunostaining shown to be evaluable with computer vision methods [9, 11, 12]. One previous study found that semi-automated analysis of Ki-67 staining with image analysis can be used for prognostic assessment of patients with breast cancer [10].

In this study, a tool for automated quantitative assessment of Ki-67 expression is presented. The tool is implemented within a previously described web-based virtual microscopy platform [23]. The IHC quantification method is evaluated by comparing the results with visual assessment of Ki-67 expression in a comprehensive series of breast cancer specimens. By linking the clinicopathological data with related tissue samples, the relationship between automated Ki-67 expression analysis and survival is assessed.

In the current report we describe the integration of an open source image analysis tool with a virtual microscopy platform. Computer determined extent of immunohistochemical staining of the extensively studied biomarker Ki-67 shows prognostic value comparable to visually assessed Ki-67 in a comprehensive series of patients with breast cancer.

The automated assessment of Ki-67 extent of staining was significantly associated with all the examined clinicopathological characteristics, including tumour size, number of positive lymph nodes, histological type and grade, oestrogen and progesterone receptor status, age at diagnosis, method of tumour detection, as well as molecular subtypes. These findings are in good agreement with previously reported results on the association between Ki-67 expression and clinicopathological factors [25, 29]. The comparison of visual and automated assessment of Ki-67 expression showed only moderate agreement. The human observer may exclude non-tumour or stromal areas in the sample more effectively than the image analysis algorithm, which may explain part of the discrepancies. Also, the image analysis algorithm can include artefacts and staining errors. On the other hand, the human interpretation can vary due to the visual evaluation being done on several separate occasions. Visual thresholds may change between scoring sessions because of altered microscopy settings or reference spots with varying stain intensity. The observed variability between visual and automated methods mostly occurred between adjacent groups, only 2 patients with negative visual score were in the automatically assessed high extent group, and none of the patients with high visual score were classified by the automated analysis into the low extent group. The two totally discrepant cases were caused by partially folded TMA spot and falsely dyed spot. In general, automated method underestimated the extent of staining in samples with high stromal content or with just a few strongly positive tumor nuclei. The main causes for too high automated scores were out-of-focus samples, debris on the glass slide or positive staining of the tumor cytoplasm.

The analysis of distant disease-free survival shows that the automated assessment of Ki-67 extent of staining is a significant predictor of outcome in breast cancer. When compared to low extent of Ki-67 staining, moderate and high extent of staining is associated with hazard ratios of 1.77 and 2.34 for distant recurrence during the follow-up period. These results are in line with previous meta-analyses, where the pooled hazard ratios for disease-free survival (DFS) associated visually determined Ki-67 overexpression have been 1.93-2.18 [19, 22].

When other variables are taken into account in the multivariate survival analysis, the automated assessment of Ki-67 extent of staining remains as a significant prognostic factor with hazard ratios of 1.62 and 1.73 for moderate and high extent of staining groups, respectively. Thus, the automated assessment of Ki-67 extent of staining is an independent predictor of patient outcome after adjustment for established clinicopathological factors. Also this is in agreement with results of meta-analyses, where a pooled DFS hazard ratio for Ki-67 overexpression in multivariate analysis was reported to be 1.76-1.84 [19, 22]. In the current study, the visually determined Ki-67 failed to reach significance in the multivariate model, possibly due to previously discussed variability in human observer assessment. The automated Ki-67 intensity assessment was of limited prognostic value. This could partly be explained by the difficulties in quantification of the diaminobenzidine staining, due to a previously described non-linear relationship between the amount of antigen and the staining intensity [30].

The strengths of this study include that we analysed a large unselected breast cancer series with long follow-up period. The software and algorithms that were utilized are open source and integrated into a virtual microscopy platform. They could be made freely available as a software service on a public web site. Examples of this approach have recently been published and represent a promising methodology for standardization of quantitative immunostaining assessment and fluorescence in situ hybridization (FISH) signal counting [17, 26]. This method could also be useful tool in routine breast cancer diagnostic pathology of prognostic and predictive factors. Even though the usage of TMA slides in this domain has shown promising results [31], these factors are mainly assessed from whole slide sections. Our approach can, however, be applied also to whole slide sections via either taking digital snapshots of regions of interest, or in case of virtual whole slides, the area to be analyzed can be selected manually.

A weakness of the algorithm proposed in our study is the need for manual adjustment of threshold levels before starting the batch analysis. A constant threshold level for the image analysis algorithm seems acceptable if similar tissue processing and staining protocols are applied throughout the whole specimen series, as in the current study. Another weakness of the current computer vision approach is that also stromal components of tissue samples were included in the automated assessment. This affects the distribution of the extent of staining as compared to studies that have excluded tumour stroma and calculated the proportion of stained cells in the tumour parenchyma only. An approach that does not exclude stroma might be acceptable for analysis of tissue microarrays that have been constructed to mainly contain tumour tissue. However, also in the TMAs the ratio of stroma to tumour epithelium can vary according to tumour grade and histological type, which can affect the extent of staining. Therefore we performed subgroup analysis according to histological grade and type, as well as adjusted for these possible confounders in a multivariate survival model. The automated assessment of Ki-67 extent of staining was a significant prognostic factor in all subgroups, except for in the group of patients with poorly differentiated tumours. This is in line with previous results showing a lack of prognostic value of Ki-67 in grade 3 tumours [32]. For analysis of individual samples and whole slide surgical tumour samples within a routine diagnostic setting, an image analysis method for excluding stroma would be needed. Examples of such method have been described in commercial systems [15, 16]. Also, a method that segments and analyses tumour nuclei only might be better suited for the Ki-67 antigen, which mainly is expressed in the cell nuclei. However, algorithms that segment tumour nuclei require an optimal nuclear counterstain, which can be hard to achieve in practice. On the other hand, a recent study showed that also cytoplasmic and membranous expression of Ki-67 is of prognostic value in breast cancer [33].

We conclude that the extent of Ki-67 staining determined by automated image analysis algorithm is an independent predictor of survival in breast cancer based on multivariate survival analysis. In univariate analysis, the automated assessment of Ki-67 extent of staining yields comparable results to the visual Ki-67 assessment. The current approach could be utilized in screening of prognostic biomarkers in large series of tissue microarrays, by integration with whole slide microscopy imaging and virtual microscopy.

The study sponsors had no involvements in the study design, in the collection, analysis and interpretation of data; in the writing of the manuscript; and in the decision to submit the manuscript for publication.