DAB2JETIA · Auto H-scorer
Continuous, per-cell IHC scoring — anchored to a pathologist.
Manual IHC scoring is the bottleneck and the variability in almost every biomarker program. A pathologist eyeballs a slide and returns two numbers — percent of tumor cells stained, and an intensity call. Those numbers move between readers, between sessions, and between sites.
DAB2JETIA measures instead of estimates. It scores every tumor cell individually, on a continuous scale — then reports the readouts your program already uses, including a classical H-score.
What makes it different
Biomarker-agnostic
One engine, not one model per stain. A new marker needs calibration, not a rebuild — so adding a target to your program is days, not a training cycle.
Continuous, not bucketed
Every cell carries a real number, not a 0/1+/2+/3+ bin. That captures the borderline cell a human must round away, and it makes small treatment effects visible.
Tumor-gated
Lymphocytes, stroma and necrosis are identified and held out, so %TC means percent of tumor cells — not percent of whatever happened to be in the field.
Pathologist-anchored
The scale is calibrated to a board-certified pathologist's reads, not to an abstract threshold. The output speaks your reviewers' language.
Auditable to the cell
Any slide-level number can be traced back to the individual cells that produced it, and reviewed cell by cell.
That correlation is best read against the variability of the reference itself. Manual IHC scoring carries well-documented inter- and intra-observer variation — the PD-L1 companion-diagnostic literature is the clearest published example, and the effect is familiar to anyone who reads these assays across sessions. An automated scale does not remove that variability from the assay; it removes it from the reading.
Compartments
Membrane is the hard problem, and it is the one solved and validated here. Measuring a thin membrane ring per cell — separating it from the nucleus and from neighboring cells, at scale, under heavy chromogen — is materially harder than measuring a nucleus or a cytoplasm. Nuclear and cytoplasmic scoring are served by the same measurement core; they are simpler compartments to isolate, not separate engines. The validation figures in this document are for membranous staining, which is the readout most companion-diagnostic-style markers report — among them PD-L1, FOLR1, TROP2, HER2, Claudin 18.2 and Nectin-4.
Four steps, one pass over the slide
Described at the level of what the engine does. The implementation — stain separation, cell delineation, cell typing and the calibration — is BCMG AI intellectual property.
Separate the chemistry
Brown chromogen and blue counterstain are separated on each slide individually, using that slide's own staining rather than fixed textbook constants. Stain intensity drifts between runs, batches and sites; measuring each slide on its own terms is what keeps the numbers comparable.
Find every cell
Every nucleus is outlined, then expanded outward to its own membrane. The measurement is taken on that membrane ring — one cell at a time, where a membranous marker actually sits.
Decide what counts
Each cell is classified and non-tumor cells are held out of the count. The denominator is the point: a percent-positive computed over lymphocytes and stroma is not the number a pathologist reports.
Score and aggregate
Measured stain density becomes a continuous 0–3 score per cell, on a scale anchored to a pathologist's reading. Those per-cell scores add up to H-score, %TC, mean intensity and heterogeneity.
Quality control is part of the measurement
Two QC outputs ship with every run, because a score without a quality statement is not decision-grade:
- Counterstain intensity — an objective Mild / Moderate / Strong grade on a fixed, absolute scale. Not re-normalized per batch, so a batch that is genuinely all-Moderate reports as all-Moderate.
- Measurable fraction — what proportion of the slide supported a valid measurement. Where chromogen density overwhelms the counterstain signal, the engine reports that and, where a negative reagent control (NRC) section is available, measures there instead.
What you get back, per slide
H-score (0–300)
The classical readout: 1×%1+ + 2×%2+ + 3×%3+ over gated tumor cells. Because scoring is already per-cell and continuous, H-score falls out directly — no second model, no separate calibration.
%TC positive
Percent of tumor cells at or above the positivity cutoff, with the equivocal fraction reported separately rather than silently rounded.
Mean intensity (0–3)
Average score across positive tumor cells, on the pathologist-anchored scale.
Heterogeneity
Both between cells and within a single cell — the half-2+/half-3+ membrane that a single categorical call cannot express.
Cell census
Tumor / immune / stroma / necrotic counts, so the denominator behind every percentage is inspectable.
QC statement
Counterstain grade and measurable fraction, per slide.
Reading the figures
Stain-intensity maps use a continuous scale from low to high:
Counterstain QC uses a separate, deliberately distinct scale so the two can never be confused when shown together:
Worked examples
Real slides from the validation cohort, spanning the expression range, scored by the current engine at its calibrated settings. Each row is one field: the input, the generated morphology view, and the analysis layers that produce the numbers. Click any panel to enlarge.
Case 1
Case 2
Case 3
Validation
Validated against board-certified pathologists — scores are drawn from routine reads, not from a single reader's annotations produced for this purpose. The working corpus spans 1,334 whole slides across multiple biomarkers and indications, with several hundred carrying pathologist intensity and %TC scores.
The table below reports a held-out evaluation: calibration was fitted on a training split and the held-out split was scored once, on a single stainer and scanner platform. Figures are regenerated from the live engine at each release, so they track the current model rather than a historical best.
| Readout | Metric | Held-out result |
|---|---|---|
| Intensity (0–3) | Mean absolute error | 0.21 |
| Systematic bias | -0.00 | |
| Correlation with pathologist | 0.75 | |
| %TC | Correlation with pathologist | 0.85 |
| Mean absolute error | 14.3 points | |
| Cohort | Calibration cohort — train / held-out split | 54 / 14 |
How to interpret these numbers
- Intensity is calibrated, not merely correlated. Held-out error of 0.21 on the 0–3 scale, with bias at -0.00, indicates no systematic tendency to read high or low against the pathologist. The offset is removed rather than averaged over.
- %TC ranks reliably; absolute agreement continues to tighten. A correlation of 0.85 means cases are ordered correctly — the property that governs screening, enrichment and stratification. Absolute agreement remains under active development and is reported as such.
- Calibration is per-marker and per-platform. Anchoring is established for a given marker, stainer and scanner combination, and re-established when any of the three changes.
Licensing & consulting engagement
DAB2JETIA is developed by a computational pathologist — board-certified in anatomic and clinical pathology, working in computer vision, digital pathology and AI model development — who also reads these assays. The scale, the gating and the failure modes are set by someone who does the task, not inferred from a label file.
Background includes Director of Digital Pathology at a national reference laboratory and Senior Pharma Pathologist, with long-standing consulting engagements — including Medical and Laboratory Directorships — for leading digital-pathology and pathology-biotechnology companies. Alongside development, the practice advises at executive level on digital pathology, the business and regulation of pathology-AI companies and their laboratories, and on standing up laboratories from scratch or converting existing ones to digital- and AI-ready operation. BCMG AI works with a small consortium of US-trained, board-certified AP/CP pathologists.
Feasibility read
Send a small labeled set. We calibrate to your marker and return an honest agreement report — including where it does not work.
Cohort scoring
Whole-slide scoring across your cohort, delivered as per-slide readouts plus per-cell data and QC statements.
Marker onboarding
Anchoring a new membranous marker to your reading convention, with the calibration evidence documented.
Reader-agreement studies
Quantify variance between your readers, and how much of it an objective scale removes.
What we need to start
- Slides — whole-slide images; a stain-free control section per run is valuable but not required.
- Labels — existing pathologist scores for a modest subset. Calibration does not need thousands.
- Your reading convention — what counts as positive in your program. We anchor to it rather than imposing ours.