He, L., Shi, R., Wang, W., Fang, G., Cai, Y., Ma, L.
EM-DINO and OmniEM are the first large-scale foundation model and unified prediction architecture for electron microscopy image analysis, providing generalized performance across diverse biological data.
Electron microscopy image analysis suffers from data heterogeneity and fragmented workflows, hindering scalable insights. Existing methods are task-specific and fail to generalize across varying conditions.
Construct EM-5M, a curated dataset of 5 million EM images, and pretrain EM-DINO, a foundation model. Using EM-DINO's multi-scale embeddings, develop OmniEM, a U-shaped architecture for unified dense prediction.
OmniEM matches the performance of EM-specific diffusion models in image restoration while reducing structural artifacts, and outperforms previous methods in 2D/3D mitochondrial and multi-class organelle segmentation. It also enables high-resolution segmentation from low-resolution inputs, applicable to legacy and high-throughput datasets.