A general method for bootstrapping dense 3D segmentations from sparse 2D annotations

A lightweight two-stage pipeline bootstraps dense 3D instance segmentations from sparse 2D labels in volumetric electron microscopy, cutting human effort by four orders of magnitude.

• Why it matters: Manual dense 3D ground truth is a major bottleneck for training accurate 3D instance segmentation models, limiting scalable volumetric bioimaging analysis.
• What they did: A 2D network predicts dense boundaries from minimal 2D labels, then a 3D network pretrained on synthetic data infers inter-section connectivity to assemble full 3D instances.
• The result: Across imaging modalities and biological structures tested, base segmentations require minutes of human effort per volume and 2D→3D pseudo ground-truth training approaches fully supervised accuracy.

AI summary of the abstract · may contain errors · not medical advice · read the paper.