Deep Learning for 3D Instance Segmentation of Densely Packed, Irregularly Shaped Nuclei

Abstract

Recent innovations in microscopy technologies has enabled rapid generation of three-dimensional (3D) live imaging data of developing biological systems, which has highlighted the need for automated extraction of information about the constituent nuclei, including their volumes, shapes, and positions. Extracting information about the nuclei requires instance segmentation: assigning each voxel to a particular nucleus or to the background. However, the instance segmentation of nuclei in 3D images is frequently challenging because of low signal-to-noise ratio, high voxel anisotropy, as well as the dense packing and irregular shapes of nuclei. Here we review state-of-the-art deep learning methods for segmenting 3D images of densely packed nuclei. We highlight how the unique challenges posed by biological data have spurred innovation in computer vision, and we give suggestions for new techniques that may improve segmentation accuracy in the future.