Phenotyping of Klf14 mouse white adipose tissue enabled by whole slide segmentation with deep neural networks

Abstract

White adipose tissue (WAT) plays a central role in metabolism, and multiple diseases and genetic mutations cause its remodeling, most notably obesity, which has reached pandemic levels. WAT is present in subcutaneous (SAT) and visceral (VAT) depots, and its main components are white adipocytes. Quantitative analysis of white adipocyte size and counts is of great interest to understand physiology and disease, due to intra- and inter-depot heterogeneity, as well as better prognosis for hypertrophy than hyperplasia, and for SAT expansion than VAT expansion. H&E histology of whole depot cross-sections provides excellent approximation of cell morphology and preserves spatial information. Previous studies have been limited to window subsampling of whole slides, and cell size analysis. In this paper, we present a deep learning pipeline that can segment overlapping white adipocytes on whole slides and filter out other cells. We also propose a statistical framework based on linear models to study WAT phenotypes with three interconnected levels (body weight BW, depot weight DW and quartile cell area). Applying it to find Klf14 phenotypes in mice using 147 whole slides of WAT H&E histology, we show sexual dimorphism, and different effects between depots, heterozygous parent of origin for the KO allele and genotype (WT vs. Het). In particular, whether variables are correlated (DW vs. BW and cell area vs. DW), and statistical differences between fitted linear models. We also find significant differences between hand-traced or window subsampling datasets and whole slide analysis. Finally, we provide heatmaps of cell size for all the slides, showing substantial spatial heterogeneity and local spatial correlations.