Deep learning for high-throughput quantification of oligodendrocyte ensheathment: a UNet architecture to extract multiple morphological parameters from individual cells

High-throughput quantification of oligodendrocyte (OL) myelination is a significant challenge that impedes the development of therapeutics to promote myelin protection or repair. Here, we established classic algorithmic and deep learning approaches to quantify myelin in a nanofiber culture system. The classic algorithm was developed by modeling general characteristics of OL ensheathments, while the deep learning neural network employed a UNet architecture that enhanced its capacity to associate ensheathed segments with individual OLs. Reliably extracting multiple morphological parameters from individual cells, without heuristic approximations, significantly improved the validity of the neural network as this mimics the high-level decision-making capacity of human researchers. Experimental validation demonstrated that the deep learning approach matched the accuracy of expert-human measurements of the length and number of myelin segments per cell. The combined use of automated imaging and analysis eliminates weeks of manual labor while reducing variability to increase reliability when screening libraries of compounds. The capacity of this technology to perform multi-parametric analysis at the level of individual cells permits the detection of nuanced cellular differences to accelerate the discovery of new insight into OL physiology.