SLiMIA - Spheroid Light Microscopy Image Atlas

Binary segmentation of 3D tumour and non-tumour spheroids in transmitted-light microscopy, spanning 9 microscopes, 48 cell lines, 8 culture media and 4 spheroid formation methods. Mirror of figshare collection 7486311.

License

CC BY 4.0 - all nine figshare data articles carry it (verified against the figshare API license field per article and against DataCite SPDX).

Note the trap: the accompanying Scientific Data paper is CC BY-NC-ND 4.0. That covers the article text and figures, not the data. Several public dataset indexes copied the article licence onto the dataset and wrongly list SLiMIA as non-redistributable.

Contents

Pairs	7,984 (image + manual mask)
Modality	transmitted light: phase-contrast (5 scopes) / brightfield (4 scopes)
Target	single foreground class - spheroid
Mask values	{0, 255} binary
Resolution	800x600 to 2592x1944
Split	none upstream - everything is train

Per-microscope

Microscope	Contrast	Pairs
Axiovert200	phase-contrast	704
Axiovert200M	phase-contrast	3,107
Cytation5	brightfield	66
IncucyteS3	phase-contrast	468
IncucyteZOOM	phase-contrast	742
LeicaDMi1	brightfield	2,311
LeicaDMi3000B	phase-contrast	439
OlympusIX05	brightfield	88
OperaPhenix	brightfield	59

Axiovert200M is additionally split by originating lab in the lab column (DeWever 2,558 / Kunz-Schughart 549) - a real batch effect worth stratifying on.

Why 7,984 and not 7,990

The atlas nominally holds 7,990 images and 7,990 masks. Six LeicaDMi1 images have no mask and six LeicaDMi1 masks have no image - genuinely different B*/T* combinations, not a name-matching failure. An image without ground truth is not evaluable, so all 12 are omitted and enumerated in unpaired.json.

Full accounting of the 15,992 upstream files: 7,984x2 paired + 6 + 6 unpaired + 11 scalebars + 1 readme = 15,992.

Ground truth

A single tier: manual foreground/background annotation in AnaSP v1.4, traced on enlarged images. There is no automated tier, no expert-corrected second pass and no multi-annotator subset, so no inter-rater agreement can be computed. Annotation was distributed across the authors plus seven named helpers, but per-image annotator identity is not recorded - annotator effects are unmodelable. The only documented QC is a two-observer image-quality screen (focus, borders, dust), not a mask review.

Empty masks are a real class, not missing annotation

107 of the 7,984 masks are entirely zero, and every one of them is IncucyteS3 (107 of that microscope's 468 frames, 22.9%). This is not corruption and not un-annotated data - it is spheroid formation kinetics, measured:

• all 107 fall at t = 0-18 h; every non-empty mask begins at 14 h or later
• 107 of the 125 IncucyteS3 frames at t <= 18 h are empty (85.6%)
• all are the same condition - MCF10A seeded at 3,000 cells
• the transition is gradual across replicates (14 h: 1 non-empty, 15 h: 2, 16 h: 4, 17 h: 5, 18 h: 6) - i.e. the spheroid progressively aggregates

So an empty mask means "no spheroid has formed yet" - a confirmed true negative. Treat it as a genuine negative, the inverse of datasets where an absent mask means "this class was not annotated".

Evaluation consequence. MedOtter's harness skips empty-reference frames rather than scoring them 0 - measured on this mirror: num_images 7984, num_processed 7877, num_skipped 107, exactly the empty set - so they do not corrupt the reported metric. They are equally not evaluated: the IncucyteS3 number below covers 361 of its 468 frames. Any pipeline computing its own metric must decide explicitly what an empty reference means. fg_fraction == 0 selects them.

Measured baseline (SAM2-large, point prompts, segmentation_mode: binary)

Whole atlas, all 7,877 scored frames: Dice 0.8651 / IoU 0.7975.

Microscope	Contrast	n scored	Dice
Cytation5	brightfield	66	0.9727
OlympusIX05	brightfield	88	0.9523
LeicaDMi1	brightfield	2,311	0.9146
IncucyteZOOM	phase-contrast	742	0.9044
Axiovert200	phase-contrast	704	0.9005
IncucyteS3	phase-contrast	361	0.8610
OperaPhenix	brightfield	59	0.8305
Axiovert200M	phase-contrast	3,107	0.8145
LeicaDMi3000B	phase-contrast	439	0.8143

Two things the split shows:

• Brightfield 0.9155 vs phase-contrast 0.8414 (+0.074). Phase-contrast puts a bright halo around the spheroid edge, which is exactly where the boundary decision is made; brightfield gives a cleaner silhouette.
• Score falls as the spheroid grows - corr(log fg_fraction, Dice) = -0.233, smallest size quintile 0.9265 down to largest 0.8060. Small spheroids are compact discs; large ones develop ragged, diffuse borders and satellite cells. This is the opposite of lesion datasets, where bigger targets score better.

No frame scored Dice 0 - there are no catastrophic failures, only boundary disagreement.

Splitting

No official partition exists. Any honest split must group on cell_line + bio_replicate: tech_replicate (T*) values are repeat acquisitions of the same well and are near-duplicates, so splitting on them leaks. The Incucyte S3 subset is a 0-374 h time series over the same wells, which leaks the same way.

Columns

image, mask, image_id, microscope, lab, contrast_mode, cell_line, medium, formation_method, seeding_density, timepoint_h, bio_replicate, bio_replicate_n, tech_replicate, tech_replicate_n, magnification, magnification_raw, width, height, fg_fraction.

magnification is normalized to lower case; magnification_raw keeps the verbatim token, because upstream is case-inconsistent (10X 4,011 vs 10x 3,168) and case-sensitive matching silently drops pairs.

Faithful-naming notes

• Complete atlas, not a subset or a re-host - author-deposited at figshare.
• Masks are stored upstream as plain .tiff, not .ome.tiff as the paper states; images are .ome.tiff. This mirror re-encodes both to PNG.
• Images are grayscale and Bio-Formats-transcoded upstream; the paper notes sources were ".jpg- or .tiff-", so an unidentified fraction passed through lossy JPEG before deposition.
• The nine per-article figshare descriptions claim "50 cell lines, 9 culture media". The collection description, the paper, and a direct count of the filenames all say 48 cell lines (47 plus the A549-MRC5 1:1 co-culture) and 8 media. The per-article blurbs are wrong.

Overlap

None with Cell Tracking Challenge, BBBC/Broad, LIVECell, NeurIPS CellSeg or the Sartorius Kaggle set. "Sartorius" in the paper is the Incucyte manufacturer, not the dataset - LIVECell is also Sartorius Incucyte phase-contrast but is 2D monolayer cells, disjoint content. The spheroids derive from the MISpheroID round-robin (Peirsman et al., Nat Methods 2021, same lab), which released a metadata knowledgebase only and never images. No cross-reference ID column exists.

Citation

Blondeel E, Peirsman A, Vermeulen S, Piccinini F, De Vuyst F, Estevao D, et al. The Spheroid Light Microscopy Image Atlas for morphometrical analysis of three-dimensional cell cultures. Scientific Data 12, 283 (2025). doi:10.1038/s41597-025-04441-x

Data DOI: 10.6084/m9.figshare.c.7486311.v1