Phenotypic Analysis of Human and Murine Endometrial Organoids using a Machine Learning Approach

Abstract

Endometrial organoids are a powerful, 3D in vitro system for studying reproductive function and disease since they are more physiologically representative than 2D cell models. However, optimal methods for the analysis of individual organoids are still developing and remain inconsistent throughout the field. To independently evaluate and optimize high-throughput analysis tools for human and mouse endometrial organoids, we first assessed several interfaces, OrganoID, OrganoSeg2, and Biodock, for their segmentation capabilities of whole-dome Z projection brightfield microscopy images. The edge detections of OrganoID and OrganoSeg2 were not sufficient to distinguish individual endometrial organoids when multiple organoids were in contact or when debris was present and thus, we did not further evaluate these interfaces for phenotypic quantitative analysis. By creating trained deep learning models on Biodock which performed reliable segmentation, we present methods to classify and quantify organoids with mixed ‘round’ or ‘abnormal’ populations, extracting critical information such as area, eccentricity, and brightness from each individual organoid. These analysis parameters were effective in organoids from human endometrial epithelium grown in complete vs. reduced medium and from genetically engineered mice. Additionally, Biodock was used to quantify live and dead cells labeled with fluorescent dyes, showing that it is also useful for quantifying high-throughput live imaging data from drug treatment experiments. Overall, deep learning models trained through Biodock accurately assessed phenotypic properties of human and murine endometrial organoids in brightfield and live-fluorescent microscopy, demonstrating that it is a promising tool for analyzing endometrial organoids to reliably detect phenotypic differences from organoid microscopy.