Poster presented at the WIN Symposium in Paris, June 28-29, 2012

Günter Schmidt, Florian Leiß, Thomas Einert, Gerd Binnig
Definiens AG, Bernhard-Wicki-Straße 5, 80636 Munich, Germany

Introduction
Since adenocarcinoma of the oesophagogastric junction is known to show human epidermal growth factor receptor 2 (HER2) overexpression we investigated the potential of IHC stained cancer tissue to provide information about disease free (DFS) and overall survival (OS) times. We compare the prognostic value of a visually assessed scoring algorithm derived from Dako HercepTest™ with results provided by data mining information from quantitative image analysis.

Methods
Three tissue microarrays (TMAs) comprising 391 cores from tissue samples of 150 patients were analysed. After IHC staining with HER2 antibody the TMAs were scanned with Zeiss MIRAX slide scanner (20x objective). Fully automated image analysis using the Definiens Cognition Network Technology® segmented and classified cells, nuclei, cytoplasm and membrane objects, and determined on a per cell basis shape, texture and color properties. Those were correlated with known DFS/OS times using a multivariate regression analysis within the R statistical software. Based on this predictive model, the patient population was divided in one group with good and one with poor prognosis by imposing a threshold on the predicted survival times. The corresponding groups obtained by the pathologist scoring were HER2 score 0, 1+, 2+ versus HER2 score 3+. Results were subsequently reproduced with Definiens Image Miner™ software.

Results
Kaplan-Meier Analysis revealed a significant (DFS: p < 2.4×10-6; OS: p < 2.5×10-7) prognostic value for the two groups generated by data mining image analysis results, whereas the visually assessed score was not significant (p>0.1).

Conclusion
Data mining quantitative image analysis may provide a more accurate evaluation of HER2 evaluation than a visual assessment of tissue samples. The quantification of HER2 overexpression by image analysis may be also highly valuable for the prediction of anti-HER2 therapy in combating this cancer type. Predictive modeling and result validation is significantly facilitated by interactive links between data points in plots and tables and the corresponding original images as provided by Definiens Image Miner.

Acknowledgements
The authors thanks Dr. Annette Feuchtinger and Prof. Axel Walch from the HelmholtzCenter Munich for image data and valuable pathology expert consultation.