Self-supervised cell segmentation in microscopy images through transformation-equivariance learning

Abstract

Cell segmentation in microscopy images is crucial for various biological analyses, yet obtaining labeled data for cell segmentation is challenging. Our research proposes a self-supervised learning method for cell segmentation, leveraging a neural network characterized by transformation-equivariance. The network’s ability to apply equivariant transformations to both input and output, combined with the symmetrical cells under specific transformations, enables efficient cell segmentation using a single image itself. We overcame constraints associated with the need for labeled data and the requirement of a solitary cell in images. Through various kinds of cell images, our method showed the performance performance comparable to traditional supervised cell segmentation approaches.