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

MitoEM 2.0 is a curated collection of eight high-difficulty 3D mitochondria instance segmentation datasets acquired using multiple volume electron microscopy (vEM) modalities, including FIB-SEM, ssSEM, and SBF-SEM. The benchmark targets biologically complex scenarios such as dense mitochondrial packing, hyperfused networks, thin-necked morphologies, and ultrastructurally ambiguous boundaries. All datasets include native-resolution volumes, expert-verified instance labels, standardized nnU-Net–compatible file structures, machine-readable descriptors (dataset.json, splits.json), and unified metadata (metadata.csv). Difficulty quantification is provided via the Dilation Collision Index (DCI) and Erosion Fragility Index (EFI). This release supports reproducible benchmarking, model development, and reuse across bioimage analysis, algorithm evaluation, and teaching.