Circular Probabilistic Based Color Processing: Applications in Digital Pathology Image Analysi

Cancer is the leading cause of death in Canada. As the only definitive diagnosis for cancer, pathology confronts more challenges with the severe social issue of aging population. Past decade has witnessed the advent of whole-slide imaging, which fosters the research of digital pathology image understanding and automatic cancer diagnosis to address challenges in pathology.;Color plays a vital role in digital pathology image analysis due to the use of chemical staining in pathology examination. However, unclear color mixing due to histological substances' co-localization and color variation among pathology images introduced by the inconsistency in pathology image preparation make reliable color-based quantitative pathology image analysis challenging. To overcome these problems, this research investigates the unique imaging model of pathology images, and introduces pathology image centered color processing algorithms based on two novel color signal treatments in the cylindrical color domain: circular probabilistic color-based pixel clustering and saturation-weighted color statistics. In the first treatment, aware of the directional nature of hue, the study innovates to model a hue distribution of an image using a circular mixture distribution, and provides a complete hue-based pixel clustering solution through maximum likelihood estimation. The second method aims to address the singularity of the HSV space in color processing. Motivated by the close relationship between saturation and hue in color perception, saturation-weighted statistics is generalized to mitigate effects of achromatic pixels on color analysis. The proposed two color treatments benefit the understanding of color content in an image.;Based on the proposed color treatments, image-dependent color estimation, blind stain decomposition for color unmixing, and a complete color normalization scheme are proposed in this dissertation. The color estimation pipeline computes the representative color of histological substances in a pathology image, building an adaptive mapping between image color and tissue content. Toward accurate stain decomposition, taking color cues obtained from the novel color treatments and the problem's physical constraint into consideration, the introduced method obtains an optimal stain separation with minimal decomposition residue in an iterative manner. In the pathology image centered color normalization scheme, by implicitly distinguishing causes of color variation, illuminant normalization and stain spectral normalization are cascaded. Extensive experimentation suggests that the introduced solutions are superior to prior arts in terms of robustness to achromatic noise, effectiveness, and capability for histological information preservation.