Color Normalization Analysis for Semantic Image Segmentation on Histopathology Images

Recent advances in semantic image segmentation have helped researchers understand an image by distinguishing different objects and understanding their relationships. Semantic image segmentation algorithms have been effectively used to identify the characteristics of complex types of tissue cells in histopathological slide images. Still, the standardization of colors has been one of the major challenges to proceeding with semantic image segmentation algorithms due to the color variation in the histopathological slide images. In this paper, we perform a two-way analysis of color normalization, evaluating four representative color normalization methods with six evaluation metrics on 19 tissue types and reducing dimensions for visualization. The experiment results show that Reinhard's color normalization outperforms other color normalization methods regarding the six evaluation metrics. Additionally, we compared the experiment results based on the color normalization methods and the tissue types using a dimensionality reduction technique. The additional experiment results demonstrate that the types of tissue images are not directly related to the color normalization results, but the dimensionality reduction technique is effective to split different color normalization methods.