Large-scale CryoET segmentation from minimal expert annotations with foundation video models

Cryo-electron tomography (CryoET) provides molecular-resolution 3D views of intact cells, but quantitative analysis at scale is limited by the time required for dense 3D annotation. Although large and diverse training datasets could eventually enable broadly generalizable segmentation models, such resources do not yet exist for the full diversity of subcellular morphology and context observed in CryoET. We introduce SABER, a segmentation framework designed to maximize the value of scarce expert annotations by replacing the need for dense volumetric supervision with sparse 2D annotations. On two CryoET datasets with ground truth, SABER matches a fully supervised baseline while reducing annotation time by up to an order of magnitude. We further demonstrate its applicability to light microscopy and volumetric electron microscopy data. By substantially lowering the burden of 3D annotation without sacrificing accuracy, SABER enables scalable generation of biologically meaningful segmentations for downstream analysis of large CryoET datasets.