The present study shows that Ki67 immunostainings can be easily evaluated using computer assisted image analysis (CAIA) on completely digitized slides. Human and CAIA evaluation are strictly correlated, although CAIA values are slightly lower compared to human evaluation.
Beside confirming the feasibility of Ki67 evaluation using an automated approach, our results underscore a few critical aspects for the use in clinical practice of the Ki67 LI evaluation. First, it is not possible to apply general cut-off values to define tumours as having a low, intermediate or high proliferative activity: cut-off values may vary as a function of the antibody used and of the method of measurement (human vs automated).
Since the introduction of the first monoclonal antibody recognizing the Ki67 molecule, the Ki67 MAb, several other MAbs have been produced and marketed without proper accurate evaluation of their different performances, which can also be remarkable [11–15]. In our routinely stained series of BC, mean Ki67 and the cut-off values to identify tertiles based on CAIA or human measurements are different for the two different antibodies used (SP6 and MM1). Although these data derive from the analysis of two different series of consecutive tumors, these results likely reflect intrinsic differences between the two antibodies, since there are no major differences in case selection and tissue fixation between the two series. A key factor to explain the different immunoreactivity between SP6 and MM1 may be the different epitopes that are targeted: SP6 targets a C-terminus region of Ki67 while MM1 targets an inner region; the possible different preservation of these epitopes after formaldehyde fixation could explain the different immunoreactivity . Another possible explanation resides in the different origin of the two antibodies, which are derived from mouse (MM1) and rabbit (SP6). Several data suggest that rabbit antibodies may indeed be more effective diagnostic tools in histopathology , in keeping with our results showing that KI67 LI obtained with rabbit SP6 MAb are constantly higher than the ones obtained with the mouse MM1 MAb.
In the present study CAIA and human evaluation, although statistically correlated, provide different Ki67 LI values, as CAIA constantly provides lower Ki67 LI values. The reason for this discrepancy could be related to a bias in identification of positive cells or the selection of tumour areas to be counted. CAIA identifies almost all cells within the selected ROI, and has a limited capacity, based on a few geometrical features, of excluding normal stromal/inflammatory cells. Thus, in this respect, CAIA is less accurate than human evaluation. However, CAIA has the advantage of measuring a much larger number of cells as compared to human operator at the microscope. This large number of evaluated cells reduces the error risk as compared to human evaluation, which is necessarily based on the count a limited number of cells, which may be not representative of the whole tumor section. Moreover human counts suffer from a bias due to the fact that positive cells are more easily identified and counted as they are more prone to capture the attention of the operator as compared to the many unstained cells.
The selection of the region of interest in tumours is another source of possible variability. We addressed this problem, by analyzing 84 cases in a blinded fashion by two different observers. The number of cells counted by the two observers, which can be assumed as an index of the extension of the selected ROIs, were statistically different. However the KI67 LI obtained by the two observers, were strictly and linearly related. This highlights the fact that CAIA is sufficiently robust to be relatively independent from the size of the ROI.
In conclusion, computer assisted can improve the accuracy and inter-observer reproducibility of Ki67 LI assessments, especially when this approach is applied to completely digitized slides. Our data showing different Ki67 LI values depending from methodology and type of antibody used suggest that it is important to standardize the methodology for Ki67 LI evaluation and that each laboratory clearly reports the antibody and the analytical procedure used. This may allow a higher quality of proliferative data collection and enhance their use in clinical practice.