Fix RAI claim.

bacterial/Croissant/metadata.json

@@ -250,7 +250,7 @@
     ]
   },
   "rai:dataLimitations": "POPSICLE Bacterial is a **multi-class compartment segmentation benchmark** for in situ bacterial cryo-electron tomograms across 8 genera (bdellovibrio, coxiella, hylemonella, hyphomonas, legionella, pseudomonas, salmonella, vibrio).\n\n**Scale.** Total scale is small (80 tomograms, 68 train / 12 test) \u2014 realistic for cryoET but unsuited to large-model scaling experiments.\n\n**Class incidence is uneven:** cytosole, intermembrane-space, and membrane appear in every run; flagellum (42 of 80) and inclusion (35 of 80) are sparser.\n\n**Out-of-scope structures.** The benchmark covers cellular compartments only; molecular complexes (ribosomes, motors, chemoreceptor arrays) inside those compartments are NOT annotated and should not be inferred from compartment masks.\n\n**Generalization** to non-bacterial cells (eukaryotic, archaeal) or to bacterial species outside the included 8 genera is not validated.\n\n**Strictly NOT recommended for:**\n\n- clinical or diagnostic use;\n- cellular structure annotation outside the 5 covered classes;\n- species-level classification;\n- structure determination;\n- instance segmentation (the masks are semantic only).",
-  "rai:dataBiases": "**Genus selection.** The 8 included bacterial genera are not a representative sample of bacterial diversity \u2014 they skew toward laboratory model organisms with relatively well-studied cellular biology and existing portal data.\n\n**Acquisition heterogeneity.** Imaging regimes vary across the 13 underlying portal datasets (acquisition microscopes, defocus, dose, tilt scheme); methods that exploit such heterogeneity may be advantaged.\n\n**Biological imbalance.** Class incidence reflects underlying biology (motors and inclusions are species-specific) rather than balanced curation.\n\n**Annotator pool.** Annotations are provided by a small expert team (9 POPSICLE / Biohub DSB authors); annotator-level disagreement is not captured in this release.\n\n**Co-hosted annotations.** The same underlying tomograms host community annotations from other CryoET Data Portal depositions (TARDIS, MemBrain-Seg, AIs/EasyMode, SAM2 EITL). This benchmark scopes strictly to deposition CZCDP-10350 to avoid label leakage, but downstream users mixing depositions must take care.",
+  "rai:dataBiases": "**Genus selection.** The 8 included bacterial genera are not a representative sample of bacterial diversity \u2014 they skew toward laboratory model organisms with relatively well-studied cellular biology and existing portal data.\n\n**Acquisition heterogeneity.** Imaging regimes vary across the 13 underlying portal datasets (acquisition microscopes, defocus, dose, tilt scheme); methods that exploit such heterogeneity may be advantaged.\n\n**Biological imbalance.** Class incidence reflects underlying biology (motors and inclusions are species-specific) rather than balanced curation.\n\n**Annotator pool.** Annotations were produced by a subset of the deposition author team (POPSICLE / Biohub DSB); the exact per-class annotator assignments are not enumerated in the deposition metadata, and annotator-level disagreement is not captured in this release.\n\n**Co-hosted annotations.** The same underlying tomograms host community annotations from other CryoET Data Portal depositions (TARDIS, MemBrain-Seg, AIs/EasyMode, SAM2 EITL). This benchmark scopes strictly to deposition CZCDP-10350 to avoid label leakage, but downstream users mixing depositions must take care.",
   "rai:personalSensitiveInformation": "**None.**\n\nThe dataset contains 3D microscopy reconstructions of biological samples (in vitro lysate or whole cells) and contains **no human-subject data**, **no personally identifiable information**, and **no demographic, geographic, health, or other sensitive personal information**. The samples are not derived from human tissue. **No IRB review was required.**",
   "rai:dataUseCases": "POPSICLE Bacterial evaluates **dense, voxel-wise multi-class semantic segmentation** of cellular compartments in cryo-electron tomograms.\n\n**Validated tasks.**\n\n1. Per-class voxel-level Dice score (paper Sec. 3.1, Sec. 5.3) computed against expert ground-truth voxel masks. 2. Cross-class mean Dice for aggregate ranking.\n\n**Construct validity** is supported by POPSICLE benchmark experiments (paper Sec. 6, Table 1) covering nnU-Net, nnU-Net-ResEnc, MedNeXt, Octopi, and SwinUNETR \u2014 showing narrow inter-method spread on dominant classes (cytosole, membrane) and consistent rank-instability on small or rare structures.\n\n**Use cases NOT validated:**\n\n- instance segmentation (the benchmark is semantic only);\n- molecular complex localization;\n- segmentation in non-bacterial cells;\n- species classification;\n- clinical or diagnostic application.",
   "rai:dataSocialImpact": "Basic-science dataset for cryo-electron tomography ML benchmarking. Enables reproducible cross-method evaluation in structural and cellular biology research; no significant societal-impact surface (no human-subject data, no clinical or diagnostic deployment, no identification or surveillance use). Released under CC0 alongside the underlying tomograms on the public CryoET Data Portal.",
@@ -306,7 +306,7 @@
         "@id": "https://www.wikidata.org/wiki/Q109719325"
       },
       "prov:label": "Expert voxel-wise segmentation (CZCDP-10350)",
-      "description": "Multi-class voxel-wise segmentation by a 9-author POPSICLE / Biohub DSB team.\n\n**Annotation schema:** one binary mask per class per run; labels reviewed for self-consistency.\n\n**Annotator authors** include Ariana Peck, Bridget Carragher, C. Braxton Owens, Dari Kimanius, Grant J. Jensen, Gus L.W. Hart, Jonathan Schwartz, Kevin Zhao, and Utz Heinrich Ermel.\n\nInter-annotator agreement scores are not currently reported with the deposition; scope is restricted to deposition **CZCDP-10350** to prevent leakage from co-hosted community annotations.",
+      "description": "Multi-class voxel-wise segmentation produced by a subset of the deposition author team (POPSICLE / Biohub DSB). The deposition metadata does not enumerate per-class annotator assignments, so this manifest does not list individual annotator names.\n\n**Annotation schema:** one binary mask per class per run; labels reviewed for self-consistency.\n\nInter-annotator agreement scores are not currently reported with the deposition; scope is restricted to deposition **CZCDP-10350** to prevent leakage from co-hosted community annotations.",
       "prov:wasAttributedTo": [
         {
           "@type": "prov:Agent",

metadata.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:321afaa6fa767a6330464335e68c1473c3233c9e965d9b4f8bd4552c33be5736
-size 25912
+oid sha256:43034a6b3a1cfdc8df0f80ac0e5cfa80a8c6e779f22a063e52b60028bb00af60
+size 25828

yeast/Croissant/metadata.json

@@ -329,7 +329,7 @@
     ]
   },
   "rai:dataLimitations": "POPSICLE Yeast is a **multi-class organelle segmentation benchmark** for *Schizosaccharomyces pombe* (fission yeast) cryo-electron tomograms.\n\n**Scale.** Total scale is very small \u2014 20 tomograms (16 train / 4 test) \u2014 making it a low-data, high-variance setting that exposes the sensitivity of segmentation methods to limited supervision. It is intentionally NOT a regime for large-model training; results report variance dominated by the train-set size.\n\n**Class incidence is highly uneven:** cytoplasm appears in 19 of 20 runs, while nucleus and nuclear-envelope appear in only 8 of 20 each (the nucleus often falls outside thin tomographic slabs).\n\n**Generalization.** Transfer to other yeast species, or to non-yeast eukaryotes, is not validated.\n\n**Class granularity.** The 6 organelle classes are coarse \u2014 the benchmark does not distinguish, e.g., early- vs. late-endosome, autophagosomes, or peroxisomes.\n\n**Strictly NOT recommended for:**\n\n- clinical or medical use;\n- cellular pathology;\n- structure determination;\n- instance-level analysis (semantic only);\n- strain or species classification.",
-  "rai:dataBiases": "**Single-source acquisition.** All 20 tomograms come from a single yeast strain set (*S. pombe*) deposited as portal datasets DS-10000 and DS-10001 by the same source lab; acquisition microscope and reconstruction protocol are uniform. Methods that exploit this homogeneity will appear to perform optimistically.\n\n**Class skew.** Distribution is uneven \u2014 cytoplasm, vesicle, and membrane-tubule are nearly always present, while nucleus and nuclear-envelope appear in only ~40% of runs.\n\n**Annotator pool.** Annotations are by the same 9-author POPSICLE / Biohub DSB team as the Bacterial deposition; annotator diversity is low.\n\n**Statistical power.** The scale (20 tomograms) is too small to support strong claims about model rankings beyond demonstrating low-data sensitivity (paper Sec. 4, Sec. 6, Table 2).",
+  "rai:dataBiases": "**Single-source acquisition.** All 20 tomograms come from a single yeast strain set (*S. pombe*) deposited as portal datasets DS-10000 and DS-10001 by the same source lab; acquisition microscope and reconstruction protocol are uniform. Methods that exploit this homogeneity will appear to perform optimistically.\n\n**Class skew.** Distribution is uneven \u2014 cytoplasm, vesicle, and membrane-tubule are nearly always present, while nucleus and nuclear-envelope appear in only ~40% of runs.\n\n**Annotator pool.** Reference annotations originate from a single upstream source (de Teresa-Trueba et al. 2023, *Nature Methods* 20:284\u2013294) and were re-deposited / curated for inclusion in POPSICLE under CZCDP-10351; the POPSICLE / Biohub DSB team did not produce new annotations for this benchmark. Annotator-level diversity is therefore minimal by construction.\n\n**Statistical power.** The scale (20 tomograms) is too small to support strong claims about model rankings beyond demonstrating low-data sensitivity (paper Sec. 4, Sec. 6, Table 2).",
   "rai:personalSensitiveInformation": "**None.**\n\nThe dataset contains 3D microscopy reconstructions of biological samples (in vitro lysate or whole cells) and contains **no human-subject data**, **no personally identifiable information**, and **no demographic, geographic, health, or other sensitive personal information**. The samples are not derived from human tissue. **No IRB review was required.**",
   "rai:dataUseCases": "POPSICLE Yeast evaluates **dense, voxel-wise multi-class organelle segmentation** in eukaryotic cryo-electron tomograms under low-data conditions.\n\n**Validated tasks.**\n\n1. Per-class voxel-level Dice score (paper Sec. 3.1, Sec. 5.3). 2. Cross-class mean Dice.\n\n**Construct validity** is supported by POPSICLE benchmark experiments (paper Sec. 6, Table 2) covering nnU-Net, nnU-Net-ResEnc, MedNeXt, Octopi, and SwinUNETR, demonstrating high variance across both models and classes \u2014 the intended signal of this low-data benchmark.\n\n**Use cases NOT validated:**\n\n- instance segmentation;\n- non-yeast eukaryotes;\n- molecular complex localization;\n- species or strain classification;\n- pathology assessment.\n\nTogether with POPSICLE Bacterial, POPSICLE Yeast is intended to expose how segmentation methods behave under low- vs. well-sampled regimes.",
   "rai:dataSocialImpact": "Basic-science dataset for cryo-electron tomography ML benchmarking. Enables reproducible cross-method evaluation in structural and cellular biology research; no significant societal-impact surface (no human-subject data, no clinical or diagnostic deployment, no identification or surveillance use). Released under CC0 alongside the underlying tomograms on the public CryoET Data Portal.",
@@ -380,8 +380,8 @@
       "prov:type": {
         "@id": "https://www.wikidata.org/wiki/Q109719325"
       },
-      "prov:label": "Expert voxel-wise segmentation (CZCDP-10351)",
-      "description": "Multi-class voxel-wise segmentation by a 9-author POPSICLE / Biohub DSB team.\n\n**Annotation schema:** one binary mask per class per run; labels reviewed for self-consistency.\n\n**Annotator authors** include Ariana Peck, Bridget Carragher, C. Braxton Owens, Dari Kimanius, Grant J. Jensen, Gus L.W. Hart, Jonathan Schwartz, Kevin Zhao, and Utz Heinrich Ermel.\n\nInter-annotator agreement scores are not currently reported with the deposition; scope is restricted to deposition **CZCDP-10351** to prevent leakage from co-hosted community annotations.",
+      "prov:label": "Annotation curation (CZCDP-10351)",
+      "description": "Reference voxel-wise organelle segmentations were originally produced by de Teresa-Trueba et al. 2023 (*Nature Methods* 20:284\u2013294) for the upstream *S. pombe* tomograms in DS-10000 / DS-10001. **No new annotations were created by the POPSICLE / Biohub DSB team for this sub-benchmark** \u2014 their role was curatorial: re-deposit under **CZCDP-10351**, canonicalize class names to the POPSICLE schema, and scope the manifest to that deposition so co-hosted community annotations do not leak in.",
       "prov:wasAttributedTo": [
         {
           "@type": "prov:Agent",