MitoVerse

A generalist 3D-EM mitochondria segmentation benchmark: many domains (modality, species, tissue, resolution) in one consistent format, with instance labels.

Code, loaders, ingestion, and PyTorchConnectomics usage: 👉 https://github.com/PytorchConnectomics/mitoverse

Each volume is one zarr store data/<dataset>/<volume>.zarr with arrays img (uint8, ZYX) and mito (instance, 0=background); per-volume metadata (voxel size, modality, species, tissue) lives in its .zattrs and in catalog.json. Semantic masks are mito > 0 — derive as needed.

3D EM Datasets with Mitochondria Annotation

* = instance labels we converted from binary segmentation. † = redistribution caveat. Shape is the average (x,y,z) per volume, ×N = number of volumes.

Microscope	Dataset	Species	Region	Avg Shape (xyz)	Resolution (nm)	# Mitos
FIB-SEM	OpenOrganelle (Xu'21), CellMap annot.	mixed	various cells/tissues	(213,204,203)×147	8×8×8	1580
FIB-SEM	└ + MitoEM2.0 chunks (Liu et al. '26)	mixed	jurkat/macro/kidney/liver/cardiac	(1024,1024,256)×10	8×8×8	2910
FIB-SEM	Müller'21 (BetaSeg)*	Mouse	Pancreas (β-cells)	(874,669,979)×7	16×16×16	1852
FIB-SEM	Mekuč'20 (UroCell)	Mouse	Urinary bladder	(256,256,256)×5	16×16×15	287
FIB-SEM	Kunduri'22	Drosophila	Gonad (spermatocyte)	(1024,1024,341)×3	16×16×16	621
FIB-SEM	Conrad'23 (MitoNet)	mixed	fly brain / C. elegans / HeLa / muscle / salivary	(546,446,466)×5	12–24 (iso)	635
ssSEM	Wei'20 (MitoEM)	Human	Temporal lobe (L2)	(4096,4096,500)	8×8×30	10552
ssSEM	Wei'20 (MitoEM)	Rat	V1 (L2/3)	(4096,4096,500)	8×8×30	5446
ssSEM	Casser'20 (Kasthuri++)*	Mouse	Somatosensory (L4/5)	(700,791,80)×2	12×12×30	265
FIB-SEM	└ Lucchi++ (Lucchi'13)*	Mouse	Hippocampus	(512,768,82)	10×10×10	70
ssSEM	Xiao'18	Mouse	Auditory cortex	(2156,2104,31)	8×8×50	468
ssSEM	Han'24	Mouse	Cerebellum (mossy fiber)	(1357,1411,331)×5	8×8×30	1144
ssTEM	Turner'22 (MICrONS, Pinky)†	Mouse	V1 (Pinky + MICrONS1)	(1203,1282,83)×16	8×8×40	1420
SBF-SEM	Jiang'25	Mouse	Brainstem	(1000,1000,100)×3	15×15×40	2057
SBF-SEM	Guay'21*	Human	Platelet	(667,736,65)×3	10×10×50	279
SBF-SEM	Haberl'18 (CDeep3M)*	Mouse	Cerebellum (brain)	(256,256,15)×2	10×10×24	247

Total: 212 volumes, 29,833 mitochondria. Volumes are organized by imaging source; benchmarks are split overlays in splits/: cellmap.json (CellMap challenge over OpenOrganelle crops) and mitoem2.0.json (MitoEM2.0 generalization suite). The OpenOrganelle *_test0/_test1 volumes are lab-annotated and belong to mitoem2.0.

MitoEM2.0 benchmark

MitoEM2.0 (Liu et al., 2026) is a multi-domain generalization benchmark of 8 sub-datasets. Official nnU-Net release: Zenodo 10.5281/zenodo.20417683. Within MitoVerse it is not duplicated — it is the split overlay splits/mitoem2.0.json over the source volumes below (muller21, openorganelle, han24, wei20, jiang25, kunduri22). EFI/DCI are per-dataset difficulty metrics from the MitoEM2.0 paper; † redistribution caveat, * converted to instances.

Name	Organism	Tissue	Cell type	Modality	Avg Shape (xyz)	Voxels	Split (tr/val/te)	# Mito	EFI/DCI	Source
ME2-Beta	Mouse	Pancreas	β-cell	FIB-SEM	(874,669,979)	3961M	3/1/3	1935	0.02/2.68	Müller'21 †
ME2-Jurkat	Human	Cell line	Jurkat	FIB-SEM	(1024,1024,256)	805M	1/1/1	372	0.00/1.62	Xu'21 *
ME2-Macro	Human	Cell line	Macrophages	FIB-SEM	(1024,1024,256)	805M	1/1/1	497	0.01/1.38	Xu'21 *
ME2-Mossy	Mouse	Cerebellum	Mossy Fiber	ssSEM	(1024,1024,500)	3333M	3/1/3	1099	0.10/4.13	Han'24
ME2-Pyra	Human	Cerebrum	Pyramidal	ssSEM	(4096,4096,500)	8388M	3/1/11	45	0.96/0.53	Wei'20
ME2-Podo	Mouse	Kidney	Podocyte	FIB-SEM	(1024,1024,256)	805M	1/1/1	1333	0.00/2.43	CellMap'24 †
ME2-Sperm	Drosophila	Gonad	Spermatocyte	FIB-SEM	(1024,1024,341)	1073M	3/1/4	775	0.01/0.96	Kunduri'22 †
ME2-Stem	Mouse	Brainstem	Mix	SBF-SEM	(1024,1024,100)	300M	1/1/1	2058	0.09/0.80	Jiang'25

License

Per-volume; see each volume's source in catalog.json. Some sources (MICrONS †, CellMap) carry attribution or redistribution caveats — check before reuse.

Acknowledgement

• Aswath et al. 2023. Segmentation in large-scale cellular electron microscopy with deep learning: A literature survey