Abstract: Pathologists study tissue morphology in order to correctly diagnose diseases such as colorectal cancer. This task can be very timeconsuming, and automated systems can greatly improve the precision and speed with which a diagnosis is established. Explainable algorithms and results are key to successful implementation of these methods into routine diagnostics in the medical field. In this paper, we propose a graph-based approach for intestinal gland classification. It leverages the high representational power of graphs for describing geometrical and topological properties of the glands. A novel, publicly available image and graph dataset is introduced based on cell segmentation of healthy and dysplastic H&E stained intestinal glands from pT1 colorectal cancer. The graphs are compared using an approximate graph edit distance and are classified using the k-nearest neighbours algorithm. With this method, we achieve a classification accuracy of 83.3%.
Keywords: Intestinal Gland Classification, Colorectal Cancer, Digital Pathology, Graph Matching, Graph Edit Distance
TL;DR: We propose a novel graph-based method for intestinal gland classification, which we evaluate on a new, publicly-available dataset.