Simultaneous Lung Cell and Nucleus Segmentation From Labelled Versus Unlabelled Human Lung DIC Images

For high-throughput and quantitative analyses of tissue microscopy images such as tissue morphology, cell type allocation and counting, simultaneous semantic and instance segmentation is needed. However, only few public data is available for developing such methods. Therefore, we provide a data set of differential interference contrast (DIC) images of human lung tissue complemented by multiplex fluorescence labelling including semantic, cell and nucleus instance annotations.We examined the structures and tasks for which fluorescence labeling is essential and compared one set of deep neural networks (DNNs) trained solely on unlabeled DIC images with another set that also utilized fluorescence labeling. Our findings indicate that while fluorescence labeling is crucial for the detection of the majority of cell nuclei, certain cells and tissue compartments can be quantified without it.Our analysis also extends currently limited knowledge about cell type composition in normal human lung tissue and contributes to further advances in quantitative tissue characterization in health and disease.