ReSCU-Nets: Recurrent U-Nets for segmentation of three-dimensional microscopy data

Abstract

Segmenting multidimensional microscopy data requires high accuracy across many images (e.g., time points or Z slices) and is thus a labor-intensive part of biological image processing pipelines. We present ReSCU-Nets, recurrent convolutional neural networks that use the segmentation results from the previous image in a sequence as a prompt to segment the current image. We demonstrate that ReSCU-Nets outperform state-of-the-art image segmentation models, including nnU-Net and the Segment Anything Model, in different segmentation tasks on time-lapse microscopy sequences. Furthermore, ReSCU-Nets enable human-in-the loop corrections that prevent propagation of segmentation errors throughout image sequences. Using ReSCU-Nets, we investigate the role of gap junctions during Drosophila embryonic wound healing. We show that pharmacological blocking of gap junctions slows down wound closure by disrupting cytoskeletal polarity and cell shape changes necessary to repair the wound. Our results demonstrate that ReSCU-Nets enable the analysis of the molecular and cellular dynamics of tissue morphogenesis from multidimensional microscopy data.