A Complete Color Normalization Approach to Histopathology Images Using Color Cues Computed From Saturation-Weighted Statistics

Goal: In digital histopathology, tasks of segmentation and disease diagnosis are achieved by quantitative analysis of image content. However, color variation in image samples makes it challenging to produce reliable results. This paper introduces a complete normalization scheme to address the problem of color variation in histopathology images jointly caused by inconsistent biopsy staining and nonstandard imaging condition. Method: Different from existing normalization methods that either address partial cause of color variation or lump them together, our method identifies causes of color variation based on a microscopic imaging model and addresses inconsistency in biopsy imaging and staining by an illuminant normalization module and a spectral normalization module, respectively. In evaluation, we use two public datasets that are representative of histopathology images commonly received in clinics to examine the proposed method from the aspects of robustness to system settings, performance consistency against achromatic pixels, and normalization effectiveness in terms of histological information preservation. Results: As the saturation weighted statistics proposed in this study generates stable and reliable color cues for stain normalization, our scheme is robust to system parameters and insensitive to image content and achromatic colors. Conclusion: Extensive experimentation suggests that our approach outperforms state-of-the-art normalization methods as the proposed method is the only approach that succeeds to preserve histological information after normalization. Significance: The proposed color normalization solution would be useful to mitigate effects of color variation in pathology images on subsequent quantitative analysis.