SiuSeven/cmfi-17-v1-0

CMFI-17: Conditioned Microscopy Field Images, v1.0

CMFI-17 is a public microscopy-field release for reproducible conditional image-generation and pseudo-instance evaluation. Its fields come from a Poterioochromonas malhamensis time-course with observed light/dark regimes and three biological replicates (A/B/C).

Scope

• Release ID: cmfi-17-v1.0
• Physical RGB/mask pairs: 4996
• Splits: train 3990, validation 503, test 503
• Conditions: 17 (0h_shared, 1h to 72h in Light and Dark)
• Image data: original 1360x1024 RGB JPEG bytes, with EXIF/metadata removed losslessly
• Mask data: paired uint16 PNG CellPose pseudo-instance masks, byte-identical to the source
• Canonical crop convention: the shared center 1024x1024 region (x0=168, y0=0), which excludes the bottom-right scale-bar annotation

What is new in v1.0

1. Reconstructed A/B/C replicate labels. The historical directory layout did not record which replicate a field came from. Every field now carries replicate_assignment_status = inferred_from_phenotype, replicate_assignment_method = phenotype_constrained_assignment_v1, biological_rep in {A, B, C} and a numeric replicate_assignment_confidence. These labels are a reconstruction, never an original record. Read the section below before using them.
2. A complete 4,996-field release. v0.1 published 4,493 fields and withheld the test split; v1.0 publishes all four thousand nine hundred and ninety-six fields including data/test.
3. Image-observable phenotype proxies for every field, with full definitions and units in metadata/feature_schema.json (density, morphology, single-cell colour and biomass).
4. Acquisition provenance. Per-field EXIF capture times are recorded in the manifest and in metadata/exif_capture.csv before the pixel files were cleaned; the published JPEGs no longer contain that metadata.
5. A replicate-stratified split that keeps the v0.1 per-condition train/validation quotas and adds the withheld test split.

Layout

canonical/fov_manifest.parquet    one row per physical field (4,996 rows)
canonical/splits.parquet         split assignment per content group
canonical/metadata.json          release provenance, counts, train-only normalization statistics
metadata/phenotype_source.tsv    every non-empty cell of the workbook TIMECOURSE sheet
metadata/phenotype_replicates.csv reported A/B/C values, verbatim, with source cell references
metadata/phenotype_summary.csv   the workbook's own Avg (Std) text, kept separate
metadata/feature_schema.json     name, formula, unit and aggregation of every image proxy
metadata/abc_assignment_report.json  the full replicate-assignment report
metadata/split_report.json       split quotas, replicate balance, v0.1 comparison
metadata/audit_report.json       the data audit gates and their outcomes
metadata/acceptance_report.json  the independent verification pass over this release
metadata/exif_capture.csv        per-field capture time and acquisition session
metadata/observation_aliases.parquet  historical Light/Dark registrations mapped to physical fields
metadata/checksums.sha256        sha256 of every file in the release
data/{train,validation,test}/images/<observation_id>.jpg
data/{train,validation,test}/masks/<observation_id>.png

Required manifest columns

Column	Meaning
observation_id	stable field ID, pm_tc_<first 16 hex of the source image sha256>
content_group_id	sha256: + SHA-256 of the decoded uint8 RGB pixels (identity boundary)
time_h, regime, condition_key	time 0/1/2/3/6/12/24/48/72 h; shared/light/dark
biological_rep	reconstructed replicate A, B or C
replicate_assignment_status	always inferred_from_phenotype
replicate_assignment_confidence	min(softmax posterior, bootstrap stability) in [0, 1]
replicate_assignment_method	always phenotype_constrained_assignment_v1
split	train, validation or test
image_path, mask_path	release-relative published paths
raw_image_sha256	SHA-256 of the source JPEG bytes
release_image_sha256	SHA-256 of the published JPEG bytes
mask_sha256	SHA-256 of the published mask bytes

The manifest also carries all 69 image proxies, per-field distances to each replicate target, per-field target errors, the second-best candidate, bootstrap stability, EXIF capture time, acquisition session, crop coordinates and the v0.1 legacy content-group ID.

A/B/C replicate reconstruction

What was done. For every condition independently: the three image proxies (valid instance count, per-cell normalised green excess at the cell level, valid foreground coverage) were robust-scaled to the condition's own median and IQR, z = (x - median) / IQR; the workbook's three replicate values of 总细胞数, 每个细胞叶绿素含量 and 干重 were expressed as positions inside their own between-replicate spread, u in [-1, +1] per dimension; a weighted distance d = sqrt(sum_k w_k (z_k - u_k)^2) to each replicate target was computed with fixed weights (1/3, 1/3, 1/3); and fields were matched to replicates by an exact minimum-cost assignment under capacity constraints that keep the A/B/C counts equal within one field. seed = 42 fixes every random choice, capacities can differ by at most one, and no field was moved by hand.

The proxy and the culture measurement live in different unit systems, so they cannot share one numeric scale: the raw spread-to-IQR ratio differs by more than three orders of magnitude across dimensions and conditions (0.02 to 65). Each dimension is therefore normalised to its own spread (amendment A1 of the frozen design), which is what equal weights can only mean; the originally frozen shared-scale rule is kept as a reported sensitivity check and assigns 60.5% of fields to the same replicate.

What the numbers say. Bootstrap stability (200 resamples per condition) and the softmax posterior are combined into replicate_assignment_confidence; 1230/4996 fields (24.6%) fall below 0.5 and are flagged low_confidence = true. The mean confidence is 0.670.

How much signal exists at all. The proxy-to-phenotype scale is not identifiable from this dataset: a culture measurement and an image proxy live in different unit systems, so the accuracy an ideal method could reach is only defined relative to an assumed coupling. Under the documented assumption that one full between-replicate phenotype spread equals one field-proxy IQR, an oracle that knows the true replicate means would reach 0.801 accuracy on average across conditions (range 0.698-0.851, chance 0.333); at half that coupling it is 0.591 and at twice it is 0.949. Treat this as an assumption display, not as a measured accuracy. What is measured: the inferred replicates concentrate in EXIF capture sessions above chance (mean majority share 0.483 versus a 0.394 capacity-preserving null, p = 0.0005), while the excluded phenotype panel shows no agreement and the workbook's own replicate orderings do not persist across time. Per-condition target separations, distances, target errors and stability values are in metadata/abc_assignment_report.json. Region-level and condition-level conclusions are much better supported than per-field replicate labels, and the acquisition-session concentration is the only independent support the reconstruction has.

What these labels are not. They are not original recorded labels, not verified replicate identities, and not evidence for cross-replicate generalization. The workbook's culture-level measurements still cannot be joined to an individual field: use biological_rep as a stratification and grouping key, and report any result as content-deduplicated field-of-view generalization. Fv/Fm, 放氧量, 脂滴数量, 类囊体面积占比 and 周质内质网占比 never entered the assignment and were read only after it was frozen; the full independent-validation results are in metadata/abc_assignment_report.json.

Split protocol

• strata: condition_key x biological_rep
• per-condition train/validation quotas are inherited from CMFI-17 v0.1; the test quota is the remainder, so every condition keeps its original share
• deterministic order within a stratum: sha256("42:<image_sha256>"), ascending
• replicate balance: iterative proportional fitting with largest-remainder rounding; the A/B/C share per split is within 0.15 percentage points of the global share
• zero content_group_id, pixel-hash, image-hash or mask-hash appears in more than one split
• normalization statistics are computed on train fields only and published in canonical/metadata.json under train_only_normalization
• validation selects models and checkpoints; the test split is evaluated once after every method is frozen
• 1237 of 4493 fields (27.5%) changed split membership relative to v0.1, by construction: v0.1 had no replicate stratification and no test split

Acquisition metadata and EXIF removal

Every source JPEG carried EXIF DateTime (Olympus DP80, camera local time). Those capture times are preserved in canonical/fov_manifest.parquet (capture_datetime, acquisition_sequence_index) and in metadata/exif_capture.csv, and are used only for provenance: the field-derived acquisition sessions (45 sessions, split at capture gaps above 120 s) are reported next to the reconstruction, but a session is not a verified replicate identity. The published JPEGs have every APP0-APP15 and COM segment removed at marker level, so no pixel is recompressed:

python -m microalgae_dynamics.data.jpeg_metadata --strip <source.jpg> <release.jpg>
python -m microalgae_dynamics.data.jpeg_metadata --verify <source.jpg> <release.jpg>

All 4996 published images were verified: decoded pixels identical to the source (pixel_equal), entropy-coded scan bytes identical (scan_equal), typically 2-3% smaller.

Verification

From the release root:

sha256sum -c metadata/checksums.sha256

This release passed the independent acceptance pass in metadata/acceptance_report.json: 4,996 fields; 3,990/503/503 split; per-condition counts equal to the workbook; A/B/C counts differing by at most one per condition; no Light/Dark duplication at 0 h; zero cross-split content duplicates; zero missing files; zero symlinks; every checksum valid.

Limitations

1. Replicate labels are inferred. No per-field A/B/C record exists anywhere in the primary sources, so no accuracy can be measured; only an informativeness bound can be reported (above).
2. Masks are CellPose pseudo-instance annotations, not human adjudicated ground truth; source instance IDs are sparse and were preserved.
3. The test split is published in v1.0 (v0.1 withheld it). Any model developed against this release inherits the responsibility for a single frozen test evaluation.
4. Acquisition order is partially recoverable from the recorded capture times and the 149-field session structure. The split controls content duplication, not temporal adjacency; report acquisition-aware analyses separately.
5. Culture and omics measurements are condition-level, not per-field. They are published as metadata/phenotype_*.csv for calibration and interpretation only.
6. The workbook is internally inconsistent in places: e.g. the identity total_chlorophyll = total_cell_count x chlorophyll_per_cell holds to 2.2% at the median but up to 84.2% for individual replicates. Reported values were preserved verbatim; no layer was overwritten. Details are in metadata/abc_assignment_report.json.

License

Data and release metadata are available under CC BY 4.0 (LICENSE_DATA.txt). Cite the release ID cmfi-17-v1.0 and the repository URL https://huggingface.co/datasets/SiuSeven/cmfi-17-v1-0 in derivative work.