BOSS: Beta-mixture Unsupervised Oligodendrocytes Segmentation System

Abstract

To develop reparative therapies for multiple sclerosis (MS), we need to better understand the physiology of loss and replacement of oligodendrocytes, the cells that make myelin and the target of damage in MS. In vivo two-photon fluorescence microscopy allows direct visualization of oligodendrocytes in transgenic mouse models, and promises a deeper understanding of the longitudinal dynamics of replacing oligodendrocytes after damage. However, the task of tracking oligodendrocytes requires extensive human effort and is especially challenging in three-dimensional images. While several models exist for automatically annotating cells in two-dimensional images, few models exist to annotate cells in three-dimensional images and even fewer are designed for tracking cells in longitudinal imaging. Furthermore, the complexity of processes and myelin formed by individual oligodendrocytes can result in the failure of algorithms that are specifically designed for tracking cell bodies alone. Here, we propose a novel beta-mixture unsupervised oligodendrocyte segmentation system (BOSS) that can segment and track oligodendrocytes in three-dimensional images over time that requires minimal human input. We evaluated the performance of the BOSS model on a set of eight images obtained longitudinally. We showed that the BOSS model can segment and track oligodendrocytes similarly to a blinded human observer. Our method offers many advantages, as it does not require fully curated data, reduces computational time, and most importantly recapitulates cell dynamic patterns observed by manually tracking oligodendrocytes. Although BOSS was developed to apply to our studies on oligodendrocytes, we anticipate that it can be modified to study four-dimensional in vivo data of any brain cell.