OmniEM: Unifying the EM Multiverse through a Large-scale Foundation Model

Room-temperature electron microscopy (EM) analysis is rapidly evolving but faces growing challenges due to massive data heterogeneity and fragmented methodologies, forming a significant bottleneck in advancing EM-based discovery. To overcome this, we present OmniEM, the first unified and extensible foundation model framework that consolidates a wide spectrum of EM tasks under one platform. At its core is EM-DINO, the first and most powerful EM-specific vision foundation model trained via large-scale self-supervised learning (SSL) on EM-5M, the largest and most comprehensive EM dataset to date, containing 5 million standardized images spanning diverse species, tissues, preparation protocols, imaging modalities, and resolutions ranging from 0.5 nm to 70 nm. Notably, this dataset was curated through a novel model-driven filtering protocol. EM-DINO demonstrates strong EM image understanding ability in semantic classification, and OmniEM extends the capacity of EM-DINO through a lightweight yet powerful U-shaped architecture, enabling impressive performance in robust mitochondrial segmentation, multi class subcellular segmentation and more. By unifying data, model, and software into a cohesive foundation, OmniEM marks a critical step toward general-purpose EM image understanding. By consolidating multiple tasks into a single framework, OmniEM streamlines EM analysis, representing a significant step toward more efficient and reusable EM workflows that accelerate scientific discovery. The data and code for this paper will be released soon.