cellgen

Simulated 3D+t fluorescence microscopy of cells, with dense ground truth on every frame. cellgen covers phenomena that existing embryo and cell-tracking data do not contain: cell death, epithelial--mesenchymal transition (EMT), cell fusion and virus-induced cytopathic effects (CPE), as well as fast shape changes such as filopodia and blebs.

This repository holds the full release: 255 sequences in 11 configurations (53,064 volumes, 74,144 nucleus tracks, 26,691 divisions), one archive per sequence. A smaller preview of 28 sequences is available as cellgen-subset.

License to be decided.

Download

Everything (about 0.9 TB compressed):

hf download Tommy1118/cellgen --repo-type dataset --local-dir cellgen

One configuration, or one sequence:

hf download Tommy1118/cellgen --repo-type dataset --local-dir cellgen --include "hl60_mitosis/*"
hf download Tommy1118/cellgen --repo-type dataset --local-dir cellgen --include "hl60_mitosis/hl60_mitosis_easy_s0.tar.zst"

Check and extract (each archive unpacks to <system>/<sequence>/):

cd cellgen
sha256sum -c SHA256SUMS --ignore-missing
for f in */*.tar.zst; do zstd -d -c "$f" | tar -xf -; done

Contents

configuration	cell line	phenomena	scene	cells	frames × Δt	voxel (µm)	sequences (controls)	varied besides tier and seed
hl60_mitosis	HL-60	mitosis, crowding	suspension	40	144 × 5 min	0.125, 0.125, 0.2	24	–
a549_filopodia	A549	filopodia (number, length, width, branching, lifetime)	single cell	1	120 × 2 min	0.125, 0.125, 1	36	–
mda231_invasion	MDA-MB-231	strong deformation, pseudopodia, invasion	3D matrix	12	120 × 5 min	0.5, 0.5, 2	24	–
pgc_bleb	zebrafish PGC	blebs at conserved volume	3D matrix	3	180 × 5 s	0.5, 0.5, 2	24	–
apoptosis_hep2	Hep-2	apoptosis (AVD, blebbing, condensation, fragmentation), 10% necrosis	monolayer	20	240 × 2 min	0.5, 0.5, 1	12 (3)	treated / control
emt_lumenoid	hiPSC	EMT, basement-membrane crossing, lumen collapse	lumenoid	90	120 × 30 min	0.5, 0.5, 2	75 (15)	induction strength (4)
emt_colony2d	hiPSC	EMT, scattering from the colony	2D colony	120	120 × 30 min	1, 1, 2	15 (3)	induction strength (4)
sars2_syncytia	U2OS-ACE2	infection, fusion into syncytia, syncytium death	monolayer	200	216 × 10 min	0.65, 0.65, 2	33 (9)	MOI (4) × TMPRSS2 (2)
vaccinia_cpe	HeLa	CPE: volume oscillation	monolayer	20	1440 × 1 min	0.5, 0.5, 1	3	–
hsv1_cpe	HeLa	CPE: condensation, surface roughening, detachment	monolayer	20	1440 × 1 min	0.5, 0.5, 1	3	–
rhinovirus_cpe	HeLa	CPE: condensation without roughening	monolayer	20	1440 × 1 min	0.5, 0.5, 1	6 (3)	treated / control (controls shared by the three HeLa configurations)

Cells and voxel size are for the medium tier.

Tiers. Relative to medium, easy has 0.6× the cell density and 2× the photon budget; hard has 1.5× the density, 0.4× the photons and twice the z step. Within a tier, sequences differ by random seed, which redraws the initial placement, the sizes, shapes, speeds and cycle phases of the cells, the timing of events, the textures and the noise. Negative controls use the same scene with the perturbation switched off.

manifest.tsv lists every sequence with its configuration, tier, condition, seed, the configuration overrides it was generated with, the result of the label consistency check, and its archive size.

Known issues

• Label consistency. A check of every label file (no gaps inside a track, valid lineage links, SEG and TRA identical) passes for 241 of the 255 sequences. The 14 others are marked fail in the label_check column of manifest.tsv; most have a one-frame gap in a single track, one (rhinovirus_cpe_hard_control_s1) has a division with a single daughter.
• Photobleaching in the virus CPE configurations. vaccinia_cpe, hsv1_cpe and rhinovirus_cpe (1440 frames) use a bleaching rate that leaves 0.3% of the signal by the last frame; the images are close to black after about 13 h. The noise-free volumes and all labels are unaffected. A re-rendered version with a lower bleaching rate may replace these files.

Layout of one sequence

<system>/<sequence>/
  raw.ome.zarr/        noisy 16-bit stack (t, c, z, y, x), 2x xy pyramid, labels/{nuclei,cells}
  clean.ome.zarr/      noise-free expected detector signal (same grid)
  phantom.ome.zarr/    pre-PSF fluorescence density
  nuc/                 Cell Tracking Challenge N-series: 01/tNNN.tif, 01_GT/{SEG,TRA}, man_track.txt
  cell/                Cell Tracking Challenge C-series (whole cells), same layout (absent for hl60_mitosis)
  STATE/cells.csv      per cell and frame: phase, death stage, EMT state, infection, syncytium,
  STATE/nuclei.csv       number of nuclei, true volume, surface area, sphericity, inside basement membrane
  STATE/events.csv     every division, death, transition, infection and fusion
  difficulty.json      CTC dataset descriptors (SNR, contrast ratio, heterogeneity, size, spacing,
                       contact, density, division rate, signal change)
  meta.json            voxel size, frame interval, PSF, track and event counts
  config.yaml          full resolved simulation configuration

Conventions. Labels follow the Cell Tracking Challenge format. A cell absorbed into a syncytium ends its cell track (event fusion in events.csv), while its nucleus keeps its own track inside the host; CTC labels cannot express merges, so events.csv is the record. A nucleus that leaves the field of view and returns starts a new track with parent 0 (01_GT/TRA/reentry_map.txt).

Simulation

Each sequence is produced in three stages: (i) an agent model updates position, shape, cell-cycle phase, mitosis and the state machine of each phenomenon; (ii) a rasterizer turns every cell and nucleus into an implicit field on a fine isotropic grid, assigning each voxel to the strongest field so that cells never overlap, with volume conserved frame to frame; (iii) an imaging model convolves with a depth-invariant PSF (vectorial confocal, widefield for HL-60), samples z planes, integrates over pixels, and applies photobleaching, background photons, Poisson shot noise, Gaussian read noise, a nonuniform camera offset and 16-bit quantization. HL-60 and A549 are calibrated against Fluo-N3DH-SIM+ 01 and Fluo-C3DH-A549 01 by median matching; the other systems use literature parameter ranges.

code/cellgen_code.tar.zst holds the simulator snapshots the sequences were generated with (_code_20261104 for all configurations, _code_20261104b for sars2_syncytia).