UC5 — Per-object CD11b morphometric cohort (control vs 4-oxo-RA): debrief + recipe

Driven live through a local Galaxy (v26.2.dev0) via the Playwright MCP + tools API, per /drive-scenario. Scenario input: UC5_IHC_issue.md (GitHub issue #28). Goal: object-level IHC morphometrics as a collection map-over, embedded on-graph, extracted to a reusable workflow.

• History: 92391ba43dd7fcf4 (“UC5 IHC CD11b morphometric cohort”)
• Page (notebook): 59c76a119581e190
• Extracted workflow: b472e2eb553fa0d1 → UC5_IHC_CD11b_DAB_extracted.ga (10 steps, 0 warnings)

TL;DR

The object-level pipeline runs end-to-end as a clean 6-element collection map-over and extracts to a 0-warning, 10-step, byte-identical-on-re-run workflow — the extraction/engineering result is solid.

The science is NOT clean (corrected after adversarial review). The faithful-to-IWC pipeline (eosin channel + per-image Otsu) gave a biologically inverted result; switching to DAB channel + a single fixed optical-density threshold flips it to the “expected” direction (4-oxo-RA: CD11b⁺ area 3.03%→0.12%, objects 242→67) — but this magnitude is confounded, not a validated effect. The treatment batch’s DAB channel is ~10× dynamic-range-compressed (per-object max DAB ≈0.12 vs control ≈1.50), so a fixed 0.1 threshold sits at treatment’s ceiling and mechanically forces it near-empty regardless of true biology. The honest signal is small and normalization-dependent — a stain-normalization re-derivation (below) collapses the apparent effect: from 26× (raw) to 3.7× (hematoxylin-mean loading control) down to ~1.2× (hematoxylin-p99 control, i.e. essentially none). One of three treated replicates (treatment_3) is a clear non-responder (it equals/exceeds controls after normalization). Conclusion: no robust drug effect can be claimed from this data+method — at best a weak, normalization-choice-dependent trend.

Stain-normalization re-derivation (added per user request, after review)

Attempted the genuine fix — real stain normalization before measuring DAB. Findings:

• Textbook Macenko is degenerate on this data. The estimated stain matrix has two near-collinear vectors (the CD11b sections are hematoxylin-dominated with sparse DAB, so the OD angular spread is too small to separate two stains). So automatic Macenko/Vahadane is not a reliable fix here.

• Hematoxylin as a loading control (the defensible normalization) — rescale each image’s DAB by the factor that matches its hematoxylin to a reference (H = nuclei = biology-invariant), then measure DAB %area > 0.1 OD. The result is highly sensitive to which H statistic sets the scale:

  normalization	control %DAB	treatment %DAB	C/T ratio
  raw (the shipped pipeline)	3.03	0.12	26.1×
  H-mean loading control	2.98	0.81	3.7×
  mean DAB optical density	—	—	1.8×
  H-p99 loading control	2.52	2.05	1.2×

• treatment_3 is a non-responder under every normalization (raw 0.32%; H-mean 2.40%; H-p99 5.03% — above the control mean). So even the modest H-mean trend rests on treatment_1/2 only.

Bottom line: the drug-effect magnitude is not identifiable from this 6-ROI, two-batch dataset with these tools — it swings from “huge” to “none” depending on an unconstrained normalization choice, and 1/3 treated replicates contradicts it. A defensible claim needs matched-acquisition data (same scanner/staining batch) or a calibration target, not post-hoc normalization. The re-derivation was done off-graph (numpy/skimage) because no reliable on-graph stain-normalization / image-math tool is installed; wiring it on-graph would encode a non-result.

What was done

1. Inputs. Built one list collection IHC ROIs (6 × 600×600 tiff) from the shipped Zenodo URLs via /api/tools/fetch — identifiers control_1..3 (Zenodo 20271100), treatment_1..3 (20157596).
2. GO/NO-GO probe (1 image). Deconvolve → split → threshold → label → features on control_1, over all 3 HED channels, to (a) confirm multi-object label maps and (b) settle the stain channel.
3. Cohort map-over (faithful-to-IWC). Ran the IWC chain over all 6: color deconvolution (rgb2hed) → split (axis=C) → extract channel index=1 → Otsu threshold → label (CCA) → per-object features.
4. Found the confound (see below); switched to the principled DAB pipeline and re-ran the cohort.
5. On-graph aggregation. collapse_dataset (add sample-id column) → datamash (per-sample count / stained-area / median morphology) → ggplot2_heatmap2 cohort figure; masks → PNG via graphicsmagick_image_convert for embedding.
6. Page. Built a notebook embedding the input collection, control-vs-treatment segmentation masks, a per-object feature table, the per-sample summary, and the cohort heatmap — every artifact on-graph.
7. Extraction. Selective extraction by job_ids (the clean DAB chain only, excluding probe/eosin cruft) → 10-step, 0-warning .ga; re-invoked on the same input for the science-identical check.

The central finding: a batch confound, and the principled fix

Symptom. With the faithful IWC pipeline (channel index=1, per-image Otsu), treatment ROIs produced more “objects” than control (1881–2790 vs 105–288) — the opposite of the expected drug effect.

Diagnosis (per-channel pixel stats on the deconvolved images). All three treatment HED channels — including hematoxylin (nuclear counterstain, present regardless of drug) — are ~10× weaker than control (e.g. H-channel max 1.8 → 0.20). Treatment originals never reach black (min RGB ~40 vs 0), i.e. compressed optical-density range → uniformly weak deconvolution → per-image Otsu on the near-blank treatment stain channel thresholds sensor noise into thousands of 1–2 px speckles. A cross-batch acquisition artifact, not biology. The object/distribution view is what exposed it (a single per-image % area would have hidden or inverted it) — which still showcases the paper’s “distributions over one number” thesis, with a methods-caution payoff.

Principled fix.

• Use the DAB channel (index=2), not the skimage-”eosin” channel (index=1). IWC’s index=1 is incidental for control and not the principled DAB readout.
• Replace per-image Otsu with a single fixed optical-density threshold (0.1) for all 6 images — an absolute “pixels above a fixed CD11b OD” measure that is batch-robust (well above background p90≈0.04, below genuine stain p99≈0.2).
• Partial validity check (NOT a full substitute for stain normalization): hematoxylin mean OD is nearly identical across arms (0.030 vs 0.031). This shows comparable average counterstain, but it is blind to the DAB dynamic-range gap that drives the effect (see caveat below) — equal mean H does not establish comparable DAB dynamic range. (These H/DAB pixel stats were computed off-graph from the raw TIFFs with tifffile, so they are not reproducible from the history as stored — a gap to close.)

Result (principled DAB pipeline):

The ~26× area / ~3.6× object reduction is the headline, but it is confounded — read the caveat. The “max DAB intensity” column (1.50 vs 0.12) is not a clean biological readout; it is the fingerprint of the ~10× cross-batch DAB compression, and a fixed 0.1 threshold applied at treatment’s ceiling re-encodes that compression as the result. Treatment “objects” are mostly noise (52–73% are 1–2 px) and their median intensity (≈0.105) sits right at the 0.1 floor. Eccentricity is similar across arms (~0.87–0.92). The defensible effect, controlling for equal hematoxylin, is the ~2× lower mean DAB OD — suggestive of a real but modest reduction in infiltrate amount; the ~26× figure overstates it.

Recipe (how to redo)

Preconditions: /drive-scenario env (backend 8080, dev 5173, no test-tools flag), ihc-morphometrics-tools.yml installed (all imgteam + devteam/column_maker + bgruening/text_processing + iuc/collection_element_identifiers

• nml/collapse_collections; plus iuc/ggplot2_heatmap2, bgruening/graphicsmagick_image_convert). Admin key from database/universe.sqlite. Map-over via the raw /api/tools batch form: {"input":{"batch":true,"values":[{"src":"hdca","id":<hdca>}]}}.

1. Upload the 6 ROIs as a named list via /api/tools/fetch (targets → hdca, elements src:url).
2. Color deconvolution map-over: ip_color_deconvolution, convtype=rgb2hed.
3. Split map-over: ip_split_image, axis=C, squeeze=false → nested list:list (6×3 channels).
4. Extract DAB channel: __EXTRACT_DATASET__ with which=by_index, index=2 (0-based), batched with map_over_type:"list" to map the outer list → flat list of 6 DAB channels.
5. Threshold map-over: ip_threshold, th_method|method_id=manual, threshold1=0.1.
6. Label map-over: ip_binary_to_labelimage (v0.7.3), setup|method=cca.
7. Features map-over (paired linked batch): ip_2d_feature_extraction, setup|mode=with-intensities, labels=<labelmaps>, setup|intensities=<DAB channels>, full feature set incl. mean_intensity.
8. Aggregate (on-graph): collapse_dataset (one_header=true, add_name=true, place_name=same_multiple) → datamash (grouping=1; count/sum/median ops) → ggplot2_heatmap2 (zscore_cond|zscore=cols, PNG).
9. Embed: graphicsmagick_image_convert masks → PNG; build the Page with history_dataset_collection_display, history_dataset_as_image, history_dataset_display directives.
10. Extract: POST /api/workflows from_history_id + curated job_ids (clean DAB chain only) + dataset_collection_ids=[1]. Re-invoke via POST /api/workflows/{id}/invocations for the identity check.

Table 1 metrics

Figures

• uc5_report_source.png — notebook source (galaxy *_display / *_as_image directives).
• uc5_report_rendered_full.png — full rendered report.
• uc5_report_rendered_masks.png — headline rendered panel: control mask dense with CD11b⁺ regions vs near-black treatment mask (embedded on-graph segmentation outputs).
• uc5_report_rendered_heatmap.png — rendered cohort heatmap (control/treatment blocks).
• uc5_workflow_graph.png — extracted workflow (10 steps; feature step shows dual Label-map + Intensity inputs).

Changes to make next time

• Skip the eosin/Otsu detour. Go straight to DAB channel (index=2) + fixed OD threshold; keep the per-image-Otsu run only as the deliberate “what goes wrong” contrast if telling the methods-caution story.
• Add explicit stain normalization (Macenko/Reinhard) if/when an imgteam tool is available, instead of relying on the hematoxylin-equivalence argument — would make the cross-batch comparison airtight and let per-image Otsu work.
• Object distribution plots. No box/violin plotter was installed (only ggplot2_heatmap2); install an iuc/ggplot2_* boxplot to add per-object area/eccentricity/intensity distributions by condition.
• Drop count_objects (0.0.5-2) — it errors on the newer giatools label format; derive objects-per-sample from the feature-table row count (done here) or via datamash count.
• Particle-size filtering. Raw CCA over-segments into 1–2 px speckle; the fixed OD threshold suppresses most of it, but a small min-area filter (or imagej2_analyze_particles_binary size>=N) would further clean the per-object distributions for morphology claims.
• More starting ROIs / true cohort. n=3 per arm is illustrative, not powered; source additional ROIs (and a real multi-core TMA for the stretch) for a stronger cohort claim.

Caveats (the central one first — added after adversarial review)

• The cross-batch confound is NOT fully fixed; the fix re-encodes it. Treatment per-object max DAB ≈ 0.12–0.18 vs control ≈ 1.50 — the same ~10× compression seen across all channels (incl. hematoxylin max) in the eosin/Otsu failure. A single fixed 0.1 threshold applied to a channel whose entire treatment range tops out near 0.12 mechanically forces treatment near-empty, independent of true CD11b biology. So the ~26× area reduction is confounded with acquisition gain, not a validated effect size. This is arm-unfair: the threshold was effectively calibrated on control (p99 of control stain) and applied to a treatment arm whose whole distribution is 10× lower.
• Not robust at n=3. treatment_3 (166 objects, 0.32% area) partially responds — within ~2× of control_3’s area. The big “26×” mean is carried by treatment_1/2 collapsing to ~0.01%.
• Object count and % area are the least robust readouts here, not the anchors — both are dominated by the gain artifact and by speckle (52–73% of treatment objects are 1–2 px; no min-area filter was applied). The most defensible comparison is hematoxylin-controlled mean DAB OD (~2×), which is far more modest.
• Proper fix (not done): real stain normalization (Macenko/Reinhard) or per-image OD rescaling so any threshold is applied on comparable dynamic ranges, then re-derive. Until then, do not assert a causal drug-effect magnitude — frame as a methods-caution vignette where neither pipeline cleanly separates biology from batch.
• n = 3 per arm — per-object distributions are large but biological replication is small.