MedOtter/3D-Platelet-EM

3D Platelet EM (platelet-em)

Dense organelle segmentation in serial block-face scanning electron microscopy (SBF-SEM) volumes of human blood platelets, from Guay et al. (Sci. Rep. 2021). Produced by the Laboratory of Cellular Imaging and Macromolecular Biophysics (LCIMB, NIH/NIBIB) with the Storrie lab (UAMS).

This mirror packages the labeled SBF-SEM volumes with 7-class semantic organelle labels (plus instance maps) as analysis-ready multipage TIFF.

Modality & specimen

• Modality: Serial block-face SEM (SBF-SEM), Gatan 3View.
• Specimen: Human blood platelets, 2 donors.
• Voxel size: ~40 x 10 x 10 nm3 (z x y x x) - strongly anisotropic (z ~= 4x in-plane).
• Task: Dense 7-class organelle semantic segmentation.

Splits

split	volume (z x y x x)	donor	source
train	50 x 800 x 800	Donor 1	canonical platelet-em
eval	24 x 800 x 800	Donor 1 (different cells)	canonical platelet-em
test	121 x 609 x 400	Donor 2 (held-out)	official reproducibility package

eval is the validation split; test is a held-out second donor for cross-subject generalization.

Label scheme (7 classes, verified against the pixel data)

index	class	original RGB color
0	background	(0, 0, 0)
1	cell	(0, 40, 255)
2	mitochondrion	(0, 212, 255)
3	alpha granule	(124, 255, 121)
4	canalicular vessel (open canalicular system)	(255, 229, 0)
5	dense granule	(255, 70, 0)
6	dense granule core	(127, 0, 0)

Encoding note. The released data uses 3 = alpha granule, 4 = canalicular vessel, which is the swap of the order given in the paper's prose. The integer encoding provided here was verified voxel-identical to the original color-coded labels (per-class voxel counts match exactly), so use the indices above.

Files

images/{train,eval,test}.tif               # uint16 grayscale SBF-SEM volume (Z pages of H x W)
labels-semantic/{train,eval,test}.tif      # uint8 semantic labels, values 0-6
labels-instance/{train,eval}-cell.tif      # per-cell instance map (original RGBA color-per-object)
labels-instance/{train,eval}-organelle.tif # per-organelle instance map (original RGBA color-per-object)

• Images are uint16. The test volume retains the raw detector range; train/eval use the canonical contrast-normalized range. Normalize per-volume (e.g. percentile scaling) before use.
• Semantic labels are integer uint8 (0-6) for all three splits.
• Instance maps are provided for train/eval only (none distributed for test); kept in the authors' original RGBA color-per-object encoding.

Provenance & notes

• Source: official LCIMB/NIBIB distribution (bio3d-vision / leapmanlab "dense-cell"). Counts match the paper (50 / 24 / 121).
• Faithful-naming caveat: this is the labeled-crops dataset; the full unlabeled acquisition volumes (250x2000x2000 and 239x2000x2000) are not part of the public release. The float64 per-voxel training-error-weighting array from the reproducibility package is not included (a method-specific training artifact, not ground truth).
• Overlap: independent specimen/lab - no shared lineage with CREMI, SNEMI3D, AxonEM, UroCell, NucMM, MitoEM, or Lucchi.
• Annotation tier: gold, expert-reviewed labels for train/eval/test. (A separate multi-rater "annotator-comparison" volume exists in the paper materials but is not part of this benchmark.)

License

US-Government work (LCIMB, NIH/NIBIB); reviewed and cleared for public release (no PII/PHI). The accompanying paper is CC BY 4.0.

Citation

Guay, M.D., Emam, Z.A.S., Anderson, A.B., Aronova, M.A., Pokrovskaya, I.D., Storrie, B., & Leapman, R.D. "Dense cellular segmentation for EM using 2D-3D neural network ensembles." Scientific Reports 11, 2561 (2021). https://doi.org/10.1038/s41598-021-81590-0