Self-supervised cell instance segmentation through transformation-equivariant representation learning

NeurIPS 2025 Workshop NeurReps Submission30 Authors

Keywords: transformation equivariance, cell instance segmentation, microscopy images

Abstract: Cell segmentation in microscopy images is a fundamental step in quantitative biological analysis, supporting tasks such as cell counting, morphological characterization, and downstream computational studies. While recent supervised deep learning methods have achieved remarkable segmentation performance, their reliance on large-scale manual annotations limits scalability across diverse imaging conditions. In this work, we present a self-supervised framework for cell instance segmentation that leverages transformation-equivariance as an inductive prior. Our method trains a neural network to produce vector fields that remain consistent under random geometric transformations. This equivariant behavior encourages structural alignment with cellular morphology, thereby enabling precise segmentation without manual labels. We demonstrate that our self-supervised method matches or exceeds pretrained supervised baselines on the LIVECell dataset and achieves strong qualitative results on additional datasets. These results highlight the potential of equivariance-driven self-supervision for label-efficient, morphology-aware segmentation. Code will be released upon acceptance.

Submission Number: 30