Meta description: AI HER2 IHC scoring targets the poorly reproducible 0-vs-1+ boundary that now gates HER2-low ADC eligibility. Here's what the evidence actually shows.

AI HER2 IHC Scoring in Breast Cancer: The HER2-0-vs-1+ Problem

Consider a breast carcinoma with faint, incomplete membrane staining in maybe a tenth of its tumor cells. Two years ago, that case got signed out as HER2-negative and nobody thought twice. Today, that same slide forces a decision — is it a 0 or a 1+? — and the answer carries real consequence, because it can determine whether a patient falls into the HER2-low category that opens the door to a class of antibody-drug conjugates (ADCs). The stain hasn't changed. The stakes have.

The Diagnostic Question

HER2 immunohistochemistry (IHC) was designed decades ago to do one thing well: identify tumors with strong, complete membrane staining (3+) that overexpress the HER2 protein, with the equivocal 2+ cases reflexed to in situ hybridization. The assay was built to separate high from not-high. It was never engineered to resolve the low end of the spectrum, where the distinction between "no staining" (0) and "faint, barely-there staining" (1+) was, for practical purposes, clinically irrelevant.

That's no longer the case. With ADCs showing activity in tumors across the HER2-low range, the 0-vs-1+ line has become a gate for therapeutic eligibility — and it sits exactly where human IHC reading is weakest. Multiple interobserver studies have shown that pathologists agree poorly at this boundary, precisely because the assay wasn't optimized to discriminate faint signal from no signal. What's notable here is that we're now asking a legacy test to do a job it was never validated for, at the one threshold where its reproducibility is worst.

That gap is the problem computational scoring aims to close.

What the Method Is and How It Works

Computational HER2 IHC scoring applies digital image analysis — increasingly deep-learning models — to whole-slide images of HER2-stained breast tissue. The software segments tumor regions, characterizes membrane staining intensity and completeness at the cell level, and outputs a HER2 category (0, 1+, 2+, 3+) using criteria meant to mirror the established scoring rubric.

There are two distinct ways to deploy such a tool, and keeping them separate matters. In standalone mode, the algorithm generates a score on its own, with no human in the loop. In AI-assisted mode, the pathologist reads the case with the algorithm's output available as decision support, then renders the final call. These are different clinical products answering different questions. Standalone performance tells you how good the machine is in isolation. AI-assist tells you whether pathologists read better with the tool than without it. A model can be excellent at one and underwhelming at the other.

The validation that matters most, then, isn't just raw accuracy — it's whether the tool moves the needle at the 0-vs-1+ boundary specifically, since that's the threshold doing the clinical work.

The Evidence

The most directly relevant data come from Krishnamurthy and colleagues, who evaluated a fully automated AI system for HER2 IHC scoring in breast cancer [1]. In standalone mode, the algorithm achieved an overall scoring accuracy of 92.1% against the reference standard [1]. On its face that's a strong number — but overall accuracy across all four categories can mask performance at the one boundary that's hard, because the easy calls (clean 3+, clean 0) inflate the aggregate.

The more telling figure is what the study reported for the 0-vs-1+ distinction under AI assistance. Pathologist agreement at that boundary rose from 69.8% without the tool to 87.4% with it [1]. A 17-point agreement gain is not trivial. It's roughly the difference between a threshold you'd distrust and one you might actually build a workflow around.

In my reading of this evidence, that 69.8% baseline is the more sobering number of the two. It quantifies, in a single study, just how shaky the unassisted human call is at exactly the line now being asked to gate therapy. Reasonable, experienced pathologists looking at the same faint membranes land on different answers nearly a third of the time. The AI-assist result suggests the tool narrows that disagreement substantially — pushing concordance into the high 80s — without removing the pathologist from the decision.

Where It Stands

A few things deserve emphasis before anyone over-reads these numbers.

First, the reference standard itself is unstable at this boundary. When the ground truth you're validating against is defined by human readers who disagree with one another 30% of the time, every accuracy and agreement metric inherits that uncertainty. An algorithm can only be as good as the consensus it's trained and measured against, and at the 0-vs-1+ line that consensus is genuinely soft. This isn't a knock on the method — it's an honest limitation of the entire enterprise until a more robust reference emerges.

Second, HER2 IHC is notoriously sensitive to pre-analytic variables — fixation time, antibody clone, staining platform, scanner. Faint 1+ staining is exactly the regime where those variables bite hardest, because small shifts in protocol can nudge a borderline case across the line. A model validated on one laboratory's slides won't necessarily hold up on another's without local checking. That's not pessimism; it's just how tissue-based assays behave.

Third, and this matters for anyone reading about regulatory status: the evidence discussed here reflects a specific study of a specific system [1]. Demonstrating an agreement gain in a validation cohort is not the same as broad, cross-laboratory, prospective validation, and it's not the same as regulatory clearance for this particular use. The 0-vs-1+ application should be understood as an area of active development, not settled practice.

What the field needs next is fairly concrete. We need a reference standard for the low end that doesn't rest solely on reader consensus — orthogonal measurement of HER2 protein or transcript could help anchor what 0 and 1+ actually mean biologically. We need multi-site, multi-platform validation that stresses these tools against the pre-analytic variation they'll meet in real laboratories. And we need prospective data showing that improved agreement at the boundary translates into more consistent, defensible eligibility determinations over time. The Krishnamurthy results are an encouraging signal that computational assistance can tighten a measurement we know to be loose [1]. Whether that signal survives contact with the messy reality of a hundred different labs is the question the next round of studies has to answer — and it's the one I'd watch most closely.

References

Krishnamurthy S, et al. Fully automated artificial intelligence–based HER2 immunohistochemistry scoring in breast cancer: standalone accuracy 92.1%, AI-assisted 0-vs-1+ agreement improvement from 69.8% to 87.4%. JCO Precis Oncol. 2024. PMID 39393036; DOI 10.1200/PO.24.00353.

AI HER2 IHC Scoring in Breast Cancer: the HER2-0-vs-1+ Problem