Quantitative assessment and classification of tissue-based biomarker expression with color content analysis.

Context: The use of computer aids has been suggested as a way to reduce interobserver variability that is known to exist in the interpretation of immunohistochemical staining in pathology. Such computer aids should be automated in their usage but also they should be trained in an automated and reproducible fashion.

Objective: To present a computer aid for the quantitative analysis of tissue-based biomarkers, based on color content analysis.

Design: The developed system incorporates an automated algorithm to allow retraining based on the color properties of different training sets. The algorithm first generates a color palette containing the colors present in a training subset. Based on the palette, color histograms are derived and are used as feature vectors to a pattern recognition system, which returns an output proportional to biomarker continuous expression or a categorical classification. The method was evaluated on a database of HER2/neu digital breast cancer slides, for which expression scores from a pathologist panel were available. The system was retrained and evaluated on different transformations of the database, including compression, blurring, and changes in illumination, to examine its robustness to different imaging conditions frequently met in digital pathology.

Results: Results showed high agreement between the results of the algorithm and the truth from the pathologist panel as well as robustness to image transformations.

Conclusions: The results of the study are encouraging for the potential of this method as a computer aid to assess biomarker expression in a consistent and reproducible manner.