MitoEM 2.0: A Benchmark for Challenging 3D Mitochondria Instance Segmentation from EM Images

We present MitoEM 2.0, a curated data resource for training and evaluating three-dimensional (3D) mitochondria instance segmentation in volume electron microscopy. The collection assembles multiscale vEM datasets (FIB-SEM, SBF-SEM, ssSEM) spanning diverse tissues and species, with expert-verified instance labels emphasizing biologically difficult scenarios, including dense mitochondrial packing, hyperfused networks, and thin filamentous connections with ambiguous boundaries. All releases include native-resolution volumes and standardized processed versions, per-volume metadata (voxel size, modality, tissue, splits), and official train/validation/test partitions to enable reproducible benchmarking. Annotations follow a consistent protocol with quality checks and instance reindexing. Data are provided in NIfTI with nnU-Net-compatible layout, alongside machine-readable split files and checksums. Baseline scripts support common training pipelines and size-stratified evaluation. By consolidating challenging volumes and harmonized labels, MitoEM 2.0 facilitates robust model development and fair comparison across methods while supporting reuse in bioimage analysis, algorithm benchmarking, and teaching.