Application of a deep learning-based image analysis and live-cell imaging system for quantifying adipogenic differentiation kinetics of adipose-derived stem/stromal cells

Quantitative methods for assessing differentiative potency of adipose-derived stem/stromal cells may lead to improved clinical application of this multipotent stem cell, by advancing our understanding of specific processes such as adipogenic differentiation. Conventional cell staining methods are used to determine the formation of adipose areas during adipogenesis as a qualitative representation of adipogenic potency. Staining methods such as oil-red-O are quantifiable using absorbance measurements, but these assays are time and material consuming. Detection methods for cell characteristics using advanced image analysis by machine learning are emerging. Here, live-cell imaging was combined with a deep learning-based detection tool to quantify the presence of adipose areas and lipid droplet formation during adipogenic differentiation of adipose-derived stem/stromal cells. Different detection masks quantified adipose area and lipid droplet formation at different time points indicating kinetics of adipogenesis and showed differences between individual donors. Whereas CEBPA and PPARG expression seems to precede the increase in adipose area and lipid droplets, it might be able to predict expression of ADIPOQ. The applied method is a proof of concept, demonstrating that deep learning methods can be used to investigate adipogenic differentiation and kinetics in vitro using specific detection masks based on algorithm produced from annotation of image data.评估脂肪来源干细胞/基质细胞分化潜能的定量方法，可通过加深对成脂分化（adipogenic differentiation）等特定过程的理解，推动这类多能干细胞的临床应用进展。传统细胞染色法常被用于检测成脂过程中脂肪区域的形成，以此作为成脂潜能的定性表征。诸如油红O（Oil-Red-O）的染色法虽可通过吸光度测量实现定量，但这类检测实验往往耗时耗材。基于机器学习的先进图像分析技术用于细胞特征检测的方案正逐渐兴起。本研究将活细胞成像（live-cell imaging）与基于深度学习的检测工具相结合，用于定量分析脂肪来源干细胞/基质细胞成脂分化过程中脂肪区域与脂滴形成的情况。不同的检测掩码（detection masks）可在不同时间点定量脂肪区域与脂滴形成情况，以此反映成脂动力学特征，并展现出不同供体间的差异。尽管CEBPA与PPARG的表达似乎先于脂肪区域及脂滴的增加，但其或可用于预测ADIPOQ的表达水平。本研究所用方法为概念验证研究，证明了基于图像数据标注生成的算法所构建的特定检测掩码，可借助深度学习方法体外探究成脂分化过程及其动力学特征。