Supervised and Unsupervised Cell-Nuclei Detection in Immunohistology

Daniel Bug, Philipp Gräbel, Friedrich Feuerhake, Eva Oswald, Julia Schüler, Dorit Merhof

Abstract: We introduce a simulation model for cell-nuclei in immunohistology that is enhanced by a cycle-consistent adversarial network to construct realistic virtual ground-truth data. This model is applied to the task of cell-nuclei detection in immunohistologically stained whole-slide images (WSI) and achieves a standalone median performance of 83.1% F1-score learning purely from synthetic annotations. We thoroughly evaluate different training scenarios with varying contributions of manual labels. It is shown that through the simulation model, the amount of required annotations can significantly be reduced without major performance losses. If only limited amounts of annotations are available, the simulation can lead to a stabilization in the detection of immune-cells.

Keywords: Immunohistology, Nuclei Detection, Virtual Ground-Truth

TL;DR: Converting from a basic simulation model to a more realistic one using CycleGAN with an additional transform consistency criterion.