A deep learning-based strategy for producing dense 3D segmentations from sparsely annotated 2D images
Analyzing volumetric microscopy data requires dense segmentation, yet training accurate machine learning models demands prohibitively expensive 3D ground-truth. To overcome this bottleneck, we present a lightweight framework that bootstraps dense 3D instance segmentations directly from sparse 2D annotations. A 2D network trained on sparse labels predicts complete boundaries on every section, and a 3D network pre-trained on synthetic data infers inter-section connectivity to build coherent 3D volumes. We validated this 2D[->]3D method across diverse datasets spanning electron, expansion, and live-cell microscopy. Strikingly, 3D models trained on these rapidly generated pseudo ground-truths achieve accuracy comparable to those trained on dense expert annotations, yielding up to a 1,000-fold reduction in human annotation time. Even accounting for downstream proofreading, total reconstruction costs drop by an order of magnitude. This approach democratizes the generation of dense 3D training data, seamlessly extending 2D foundation models into the third dimension.