Abstract Background: Immunohistochemical Ki67 evaluation reflects proliferative activity and is one of most important prognostic/predictive markers of breast cancer. However, standardized and efficient methodologies to accurately and reproducibly measure the Ki67 expression are still to come. Besides tissue processing, sampling, intra-tumour variability, and many other aspects to be considered, key element of the methodology remains accurate enumeration of Ki67-labelling index (LI). We aimed to develop a methodology to estimate and improve accuracy of automated image analysis (IA) approach. Methods: Tissue microarrays (1 mm diameter spot per patient, n=164) from invasive ductal breast carcinoma, stained for Ki67 and digitized by Aperio XT scanner, were used for the study. Reference values (RV) were obtained by counting the LI using stereological frame overlaid on a spot image. To test the degree of inter-observer variation in establishing the RV, the frame counts were performed by 3 observers independently in a subset (n=30) of the TMA images. IA was performed with Aperio Genie/Nuclear algorithms enabling automated selection of tumour tissue. Accuracy of the IA compared to the RV was estimated based on ANOVA, correlation and regression analyses performed with SAS 9.3. Agreement between individual measurements was also estimated based on 95% confidence intervals calculated from the RV according to stereology rules. Several iterations of the IA with adjusted algorithm settings were performed to improve the accuracy. Highly automated calibration cycles were enabled by developing software to integrate processes of the image and statistical analyses. Visual evaluation for the LI on the same images was performed by 3 pathologists (P1, P2, P3). Results: Inter-observer variation between 3 independent frame counts (n=30) was negligible by ANOVA (respectively, mean RV were 28.5, 28.6 and 29.9%) with correlation coefficients 0.97 and above. RV correlated strongly with IA (r=0.95) and P1, P2, P3 (r=0.86, r=0.90, r=0.92, respectively), p<0.0001. ANOVA revealed no significant pairwise differences of the LI means of RV(40%) versus IA(37%), P2(43%), or P3(44%); however, the IA versus P2, P3 differed, and P1(24%) was significantly lower compared to all other measurements (p<0.05). Regression analysis to predict the RV revealed best performance for the IA results with R-square of 0.90 compared to 0.74, 0.82, and 0.85 of P1, P2, and P3, respectively. The IA results reported above were achieved by a third iteration of the IA calibration, while the results of the initial two IA settings revealed a significant bias to lower values by ANOVA and lower R-square values (0.86 and 0.87) by regression analyses, when compared to the RV. Similarly, correlation coefficients and agreement between individual IA measurements and RV improved during the calibration process. Conclusion: Our experiments provide sound and efficient methodology to achieve accurate immunohistochemical Ki67 enumeration by IA, enabled by proper validation and calibration of the measurement against RV obtained by stereological frame counts. Citation Format: Arvydas Laurinavicius, Benoit Plancoulaine, Aida Laurinaviciene, Paulette Herlin, Raimundas Meskauskas, Indra Baltrusaityte, Justinas Besusparis, Nicolas Elie, Philippe Belhomme, Yasir Iqbal, Catherine Bor-Angelier. A methodology to ensure and improve accuracy of Ki67 digital immunohistochemistry analysis in breast cancer tissue. [abstract]. In: Proceedings of the AACR Special Conference on Advances in Breast Cancer Research: Genetics, Biology, and Clinical Applications; Oct 3-6, 2013; San Diego, CA. Philadelphia (PA): AACR; Mol Cancer Res 2013;11(10 Suppl):Abstract nr B116.