When the Algorithm Meets the Faintest Blush: Computational HER2 Analysis and the Ultralow Problem

There's a particular kind of humility that sets in when you realize the machine is stuck in exactly the same spot you are. That's roughly where computational HER2 image analysis finds itself in 2025 — impressively good at counting membrane staining, and still stumped by the same faint, ambiguous blush of chromogen that divides pathologists reading the same slide.

The development

The immediate trigger for this reckoning is a body of work asking a pointed question: can artificial intelligence make the HER2-low and HER2-ultralow call reproducible where human eyes cannot?

A June 2025 investigational study put the tools to a direct test. Researchers examined 47 advanced breast carcinoma cases across matched core biopsies, surgical resections, and distant metastases, scoring HER2 status by three methods in parallel — conventional light microscopy, digital pathology on whole-slide images, and AI-assisted image analysis [1]. The design is unusually demanding because it tracks the same tumor through space and time, which is precisely where classification instability tends to hide.

Alongside that primary data, a January 2025 review synthesized the broader literature and reached a candid conclusion: AI tools are "stumped" by the HER2-low boundary in much the way pathologists are [2]. And a forward-looking 2026 state-of-the-field review, published in mid-2025, situates all of this within the wider trajectory of digital and computational pathology — adoption standards, clinical integration, and the specific gap between AI tools that carry regulatory clearance and their validated performance at the HER2-low/ultralow threshold [3]. Taken together, these aren't announcements of a solved problem. They're a shared acknowledgment that the hardest part of HER2 scoring may be irreducible with current approaches.

The diagnostic problem

To understand why this matters, you have to appreciate how much diagnostic weight now rests on a distinction the original HER2 assay was never designed to make.

HER2 (ERBB2) immunohistochemistry in breast cancer was built to sort tumors into a clinically actionable dichotomy: positive (3+, or 2+ with amplification confirmed by in situ hybridization) versus negative. Everything below the amplification line — IHC 0, 1+, and non-amplified 2+ — was lumped together as "HER2-negative" because none of it predicted benefit from the classic HER2-targeted agents. The scoring anchors, the training, the proficiency testing all evolved around that binary.

The arrival of antibody–drug conjugates changed the stakes. Suddenly the tumors formerly dismissed as negative were resolved into meaningful strata — HER2-low (IHC 1+, or 2+ without amplification) and, more recently, HER2-ultralow (a faint, incomplete membrane staining in more than 10% of cells, sitting just above a true 0). This matters for eligibility considerations around an entire drug class rather than any single, patient-directed regimen. But the assay is being asked to draw a line it wasn't calibrated to draw: the boundary between 0 and 1+, and between 0 and ultralow, hinges on faint, incomplete, sometimes barely perceptible membrane staining.

That's an unforgiving place to demand reproducibility. Inter-observer agreement, robust at the 3+ end, degrades sharply at the low end. Two competent pathologists can look at the same faint staining and land on different sides of a boundary that now carries therapeutic weight. This is the reproducibility problem computational analysis was supposed to help solve — and it's the substrate on which all three developments sit.

The evidence

Here's where the evidence gets genuinely interesting, and genuinely sobering.

The 47-case study demonstrates two things at once. First, HER2-low and HER2-ultralow status is dynamic — it converts across time points and tissue types, so a core biopsy, its matched resection, and a distant metastasis from the same patient can disagree with one another [1]. That's a biological signal, not a technical artifact, and AI-assisted digital pathology captured these changes alongside conventional microscopy. Second — and this is the crucial caveat — the study found that AI does not resolve the classification instability inherent to the scoring boundary itself [1]. It can measure the moving target faithfully. It cannot make the target stop moving.

The January 2025 review reinforces the point from the technical side. Computational image analysis genuinely improves quantitative membrane scoring — it can count stained cells and grade completeness more consistently than an eye estimating percentages under time pressure. But that quantitative gain doesn't translate into eliminating the core reproducibility problem at the IHC 0-versus-1+ threshold [2]. The evidence suggests the bottleneck isn't the algorithm's measurement precision; it's the absence of standardized training datasets and consensus scoring anchors to teach the model where the line actually falls. If the ground truth is contested, a model trained on contested labels inherits the contest.

The 2026 state-of-the-field review adds the systems-level view: AI-driven HER2 scoring has advanced, and clinical integration milestones — including HER2-low use cases — are real, but a gap persists between tools that carry regulatory marking and their demonstrated performance at exactly the clinically critical boundary [3].

Where it stands

So what's the honest status?

Every development discussed here is investigational in the sense that matters most — none has been shown to resolve the ultralow reproducibility problem, and the reviews are explicit that a gap remains between CE-marked or FDA-cleared AI HER2 tools and validated performance at the low/ultralow threshold [3]. Clearance for a HER2 scoring workflow is not the same as validated reproducibility at the 0-versus-ultralow call. Readers should hold those two facts apart.

The open questions are concrete. Can consensus scoring anchors be established for ultralow, and can training datasets be standardized enough that models trained in one lab generalize to another [2]? How should the field handle the biological reality that HER2-low status genuinely changes across a patient's disease course, so that no single measurement — human or machine — is the whole story [1]? And what validation bar should regulators and laboratories demand specifically at the boundary that carries therapeutic consequence, rather than at aggregate accuracy that's dominated by easy 0s and easy 3+s?

The measured read is this: computational HER2 analysis is a real advance in consistency of measurement, and a real disappointment as a solution to the boundary problem — because the boundary problem was never only about measurement. It was about where we've agreed to draw the line, and we haven't fully agreed yet. The algorithm, it turns out, is waiting on us.

---

References

1. Not extractable from snippet — see PMC record. Changes in HER2low and HER2-ultralow status in 47 advanced breast carcinoma core biopsies, matching surgical specimens, and their distant metastases assessed by conventional light microscopy, digital pathology, and artificial intelligence. PMC / peer-reviewed journal, 2025. PMCID: PMC12396990. https://pmc.ncbi.nlm.nih.gov/articles/PMC12396990

2. Not extractable from snippet — see MDPI record. HER2-Low Breast Cancer at the Interface of Pathology and Technology: Toward Precision Management. Biomedicines (MDPI), Vol. 14, No. 1, Article 49, 2025. DOI: 10.3390/biomedicines14010049. https://www.mdpi.com/2227-9059/14/1/49

3. Not extractable from snippet — see PMC record. What's new in digital and computational pathology 2026: advances in adoption, standards, AI technologies, and clinical integration. PMC / peer-reviewed journal, 2025. PMCID: PMC13183467. https://pmc.ncbi.nlm.nih.gov/articles/PMC13183467