Hierarchical Graphical Models and Graph Neural Network Explainability in Digital Pathology – Orcun Göksel

Cancer diagnosis, prognosis, and therapy response predictions from tissue specimens highly depend on the phenotype and topological distribution of constituting histological entities. Thus, adequate tissue representations for encoding histological entities is imperative for computer aided cancer patient care. We treat the tissue as a hierarchical composition of multiple types of histological entities from fine to coarse level, modeling histological entities from cell to tissue level with their intra- and inter-entity level interactions. Evaluations are shown on BReAst Carcinoma Subtyping (BRACS) and prostate cancer datasets.

Furthermore, for explainability over such entity-based graphical models to go beyond conventional pixel-level explanations, we propose graph explainers. To show their accessibility to pathologists as meaningful explainers in a standardized and quantifiable fashion, we propose a set of novel quantitative metrics based on statistics of class separability using pathologically measurable concepts. We employ the proposed metrics to evaluate three types of graph explainers, namely the layer-wise relevance propagation, gradient-based saliency, and graph pruning approaches.

About Orcun Göksel