HACT-Net: A Hierarchical Cell-to-Tissue Graph Neural Network for Histopathological Image Classification

Cancer diagnosis, prognosis, and therapeutic response prediction are heavily\ninfluenced by the relationship between the histopathological structures and the\nfunction of the tissue. Recent approaches acknowledging the structure-function\nrelationship, have linked the structural and spatial patterns of cell\norganization in tissue via cell-graphs to tumor grades. Though cell\norganization is imperative, it is insufficient to entirely represent the\nhistopathological structure. We propose a novel hierarchical\ncell-to-tissue-graph (HACT) representation to improve the structural depiction\nof the tissue. It consists of a low-level cell-graph, capturing cell morphology\nand interactions, a high-level tissue-graph, capturing morphology and spatial\ndistribution of tissue parts, and cells-to-tissue hierarchies, encoding the\nrelative spatial distribution of the cells with respect to the tissue\ndistribution. Further, a hierarchical graph neural network (HACT-Net) is\nproposed to efficiently map the HACT representations to histopathological\nbreast cancer subtypes. We assess the methodology on a large set of annotated\ntissue regions of interest from H\\&E stained breast carcinoma whole-slides.\nUpon evaluation, the proposed method outperformed recent convolutional neural\nnetwork and graph neural network approaches for breast cancer multi-class\nsubtyping. The proposed entity-based topological analysis is more inline with\nthe pathological diagnostic procedure of the tissue. It provides more command\nover the tissue modelling, therefore encourages the further inclusion of\npathological priors into task-specific tissue representation.\n