Appearance
The Benchmark Data
Seven published, fully scored benchmark pairs on Zenodo: DOI 10.5281/zenodo.22084921, CC-BY-4.0.
The Volumes
Six development volumes spanning FIB-SEM and cryo-ET, plus one held-out volume:
| Volume | Modality | nm/px | Shape (z×y×x) | Classes | Source |
|---|---|---|---|---|---|
jurkat1 | FIB-SEM | 4.0 | 500×250×500 | 21 | OpenOrganelle jrc_jurkat-1 |
macrophage2 | FIB-SEM | 4.0 | 256×256×256 | 13 | OpenOrganelle jrc_macrophage-2 |
Brno | cryo-ET | 0.784 | 120×657×652 | 14 | CZ CryoET DP 10301 (EMPIAR-11756) |
C.eleg | cryo-ET | 1.348 | 57×428×513 | 6 | CZ CryoET DP 10477 (EMPIAR-12049) |
Enco | cryo-ET | 2.638 | 40×202×321 | 9 | CZ CryoET DP 10438 |
S.pombe | cryo-ET | 1.349 | 33×400×787 | 8 | CZ CryoET DP 10000 (EMPIAR-10988) |
heldout (25jul29a_Position_3) | cryo-ET | 1.4985 | — | 5 | CZ CryoET DP 10456 |
The paper's ML tier ran on macrophage2, Brno, jurkat1.
What Each Development Volume Contains
| File | Content |
|---|---|
clean.tif / noisy.tif | The canonical frozen-config pair (float32, full depth) — behind every quantitative result |
clean_v2.tif / noisy_v2.tif | A second, visually-tuned configuration (config-sensitivity check; the gallery figure) |
labels.tif | The segmentation the phantom was rendered from |
config_frozen.json / config_visual.json | The exact render configs — every stack is reproducible |
noise_crop.npy + noise_crop.noisefit.json | The background crop and its fitted v4 noise model |
eval_window.json | The scoring region (z_block, y_window, x_window) + nm_per_px, deployed sd, structure axis |
The Held-Out Volume
heldout/ differs by design: a single frozen config (floor-matched sd), a noise_crop.tif, crop_provenance.json, an evaluation window spanning the whole working volume, and labels combining deposited masks (membrane, lysosome, mitochondrion) with rasterized oriented pick points (ribosomes, microtubules). Its run5_records/ folder preserves the pre-registration, the even/odd half-sum reference volumes (Noise2Noise argument), and every score file of the held-out study — the τ +0.67 vs −0.71 result is fully traceable.
Scores
scores/ carries the published validation runs: T1/T2 score files for the canonical (run 2) and visually-tuned (run 3) configurations, the production noise-refit report, and installed_versions.txt (careamics 0.3.2, torch 2.9.1+cu128, bm3d 4.0.3).
Licences & Attribution
| Item | Licence |
|---|---|
| The deposit (pairs, configs, crops, scores) | CC-BY-4.0 |
| PhantEM code | BSD-3-Clause |
| cryo-ET source data (CZ CryoET Data Portal) | CC0 |
| OpenOrganelle FIB-SEM source data | CC-BY-4.0 |
Source accessions, for citing the underlying acquisitions: CryoET Data Portal datasets 10301 (run RN-14070), 10477 (RN-33585), 10438 (RN-15937), 10000 (RN-242), 10456 (run 25jul29a_Position_3); OpenOrganelle jrc_jurkat-1 (DOI 10.25378/janelia.13114259) and jrc_macrophage-2 (DOI 10.25378/janelia.13117745).
Citing
Please cite the PhantEM paper (preprint link coming soon) and the data DOI 10.5281/zenodo.22084921. Code: github.com/lucasfortune/phantem, results at tag v1.1.