TODO.md

MitoVerse — Open Questions / TODO

Unsolved items as of the current build (212 volumes, 13 datasets). Grouped by theme.

Data integrity

• kunduri22 gt not connected. train01/train02 are separate stores; an instance crossing that z-cut gets a different id in each. Ingested count is 621 (218+172+231) vs the paper's 775 — likely this split + double/under-count. Connect the gt across train01/train02 (and decide whether to then merge into one 512z volume).
• ME2-Pyra # Mito = 45? The MitoEM2.0 table lists ME2-Pyra at 45 mitochondria, but wei20/mitoEM-H has 10,552. Confirm the ME2-Pyra figure (is it per held-out crop, or an error?).
• wei20 test withheld. Only train+val (500z) ingested for MitoEM-H/R, to keep the grand-challenge test hidden. Confirm this stays the policy for any public release.

Missing metadata

• turner22 voxel = 8×8×40 (MICrONS default) — verify the Pinky-specific resolution; the 5 microns1_vol* crops may differ.
• conrad23 voxel/species — DONE (filled per-volume from the MitoNet paper; FIB-SEM isotropic 12–24 nm).
• Per-volume license / redistribution rights — especially unpublished OpenOrganelle crops, MICrONS (†), and internal kunduri22. Clear before any HuggingFace push.

Not-yet-ingested data

• MICrONS1 full volume (mito/lichtman/microns1_*): 16 tiles of (750,1024,1024). Decide: stitch the 4×4 grid into one volume and add to turner22, or skip. (Only the 5 small microns1_vol* crops are in.)
• Peng's other 3D sets: cerebellum (4 quadrant vols, instance), wilson19, zhu22 — voxel/metadata TBD. Boston3Mouse is image-only (no labels) → skip unless labels appear.
• OME-Zarr refined labels: only ME2-Sperm was taken from MitoEM2.0_OMEZarr. The other 7 ME2 OME-Zarr sub-datasets may carry refined labels vs the MitoLE-sourced volumes — decide whether to re-ingest from the OME-Zarr (needs the zarr-v3 path).

Candidate external datasets to acquire (mito · 3D)

Sourced from the torch-em EM dataset tracker (full sheet cached at <pytc>/lib/mitoverse/tmp/em_datasets.csv). Filtered to mitochondria + volumetric + usable labels.

Ready — instance mito, 3D, not yet in MitoVerse

• VNC (Gerhard'13) — ssTEM, Drosophila larval ventral nerve cord; instance mito. Small, easy add. Loader: https://github.com/constantinpape/torch-em/blob/main/torch_em/data/datasets/electron_microscopy/vnc.py (NB: different from our COSEM jrc_fly-vnc-1.)
• MBLiver (Nat Commun 2024) — FIB-SEM, mouse liver; instance mito. ~2.1 TB native → take a 16 nm ROI / downsample. https://doi.org/10.6019/EMPIAR-12017 · paper https://www.nature.com/articles/s41467-024-48272-7
• MICrONS minnie (cortical mm³) — ssTEM, mouse cortex; instance mito (+ neurites). Complements our Pinky (turner22). https://www.microns-explorer.org/cortical-mm3#voxel-segmentation · https://zenodo.org/records/5760218
• MitoNet_OpenOrganelle_Mouse_Kidney — auto MitoNet instance mito, not proofread → low priority. https://figshare.com/articles/dataset/MitoNet_automatic_instance_segmentation_of_mitochondria_in_the_OpenOrganelle_Mouse_Kidney_dataset/20749729/2

Needs binary→instance conversion (semantic mito source)

• ASEM — FIB-SEM whole-cell; semantic mito + golgi/ER/nucleus → cc3d to instances. https://doi.org/10.1083/jcb.202208005 · loader https://github.com/constantinpape/torch-em/blob/main/torch_em/data/datasets/electron_microscopy/asem.py
• DeepContact — semantic mito/ER/lipid (figshare server was flaky). https://figshare.com/articles/dataset/DeepContact_Training_Data/19898404/1
• Human Organoids — FIB-SEM, semantic organelles incl. mito; EMPIAR-11380. https://www.ebi.ac.uk/empiar/EMPIAR-11380/
• ProbTEM — semantic mito but 2D TEM → skip for the 3D benchmark. https://yoonlab.unist.ac.kr/index.php/research/mitochondria-tem-dataset/

Extend what we already have

• CellMap — more annotated jrc_* crops (being extended upstream). https://doi.org/10.25378/janelia.c.7456966 · loader https://github.com/constantinpape/torch-em/blob/main/torch_em/data/datasets/electron_microscopy/cellmap.py
• OpenOrganelle — additional jrc_* volumes with mito (e.g. mb-liver, more mus-liver-zon/kidney). https://openorganelle.janelia.org/datasets

Already in MitoVerse (no download)

Kasthuri & Lucchi (casser20) · MitoEM (wei20) · UroCell (mekuc20) · BetaSeg (muller21) · DenseCell/Guay (guay21) · Haberl (haberl18) · Xiao'18 · Jiang'25 · CEM/MitoNet-3D-benchmark (conrad23) · OpenOrganelle/CellMap subset (openorganelle) · Pinky/MICrONS (turner22).

Benchmark splits

• cellmap.json: all 147 crops currently marked train. Apply the official CellMap challenge train/val/test split.
• mitoem.json: not written yet — classic MitoEM (wei20) split (region-based).
• Within-volume regions (PyTC feature). ME2-Pyra (wei20 crops) and classic MitoEM splits need PyTC to support coordinate regions as train/val/test within one volume. Until then they reference whole volumes.
• Self-annotated OpenOrganelle volumes not in any split (e.g. some cardiac/mus-liver test crops) have ambiguous annotation provenance — decide where they belong.

Tooling / publishing

• zarr v3 env. OME-Zarr ingestion currently uses a throwaway pip venv. Stand up a persistent conda env (conda create -n zarrv3 -c conda-forge "zarr>=3" ...) for repeatable v3 reads.
• HuggingFace push (pending, human auth): (a) MitoVerse data + splits + card; (b) deprecation banners on pytc/MitoEM and pytc/MitoEM2.0 (committed locally at weidf/lib/hf_legacy_cards/). Also fix the malformed YAML frontmatter on the pytc/MitoEM card (line 1 sits above the opening ---).

Definitions

• EFI / DCI difficulty metrics in the MitoEM2.0 table are quoted but not defined here — add their definitions (or a paper reference) so the numbers are interpretable.