Unsupervised Domain Adaptation Using Temporal Association for Segmentation and Its Application to C. elegans Time-Lapse Images

Abstract

AbstractWith the widespread use of live-cell imaging technologies for observing the dynamics of live cells under a microscope, there is a need for semantic image segmentation to quantitatively detect structural changes of the targeting objects in the various resulting time-lapse images. Although supervised learning of convolutional neural networks such as U-Net is one way to achieve segmentation, pixel-level labeling for images is a time-consuming task and requires domain knowledge. Unsupervised domain adaptation (UDA) methods have been proposed to transfer the learning knowledge gained from labeled datasets (i.e., the source domain) to unlabeled datasets (i.e., the target domain). However, to date, there has been no reported application of a UDA method to consider the temporal association of features with the aim of using semantic segmentation for time-lapse microscopy images. This paper proposes a new UDA method that considers the class-specific feature distribution in both the source and target domains as well as the temporal association of features. We also present manually created annotations of true nuclear regions for two kinds of open time-lapse image datasets of Caenorhabditis elegans early embryos. Using the annotated datasets, we demonstrate that the accuracy of the proposed method is more than 7% higher than that of other UDA methods with static images and videos as inputs. The code and annotations are available at https://github.com/tohsato-lab/T-MCD.KeywordsDomain adaptationVideo semantic segmentationBiological benchmark datasets Caenorhabditis elegans