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

MitoEM 2.0 is a difficulty-aware extension of the original MitoEM benchmark for training and evaluating three-dimensional mitochondria instance segmentation methods in volume electron microscopy. The dataset aggregates multiscale vEM volumes from FIB-SEM, SBF-SEM, and ssSEM imaging across diverse tissues and species. It includes expert-verified instance labels covering biologically challenging cases such as dense mitochondrial packing, hyperfused networks, and thin filamentous connections with ambiguous boundaries. This release provides native-resolution image volumes, standardized processed versions, per-volume metadata, official train/validation/test splits, and quality-controlled annotations. Data are distributed in NIfTI format with an nnU-Net-compatible layout, together with OME-Zarr 0.4 files for cloud-aware and lazy-loading workflows. The release also includes split files, checksums, and baseline scripts for common training pipelines, size-stratified evaluation, and instance reindexing. By consolidating challenging volumes and harmonized labels, MitoEM 2.0 supports reproducible benchmarking, robust model development, and fair comparison across mitochondria segmentation methods.