An Image-Derived Risk Score Enters the HR+/HER2− Decision, and the Workflow Gets Interesting

The development

For years, the question of whether a patient with HR+/HER2− early breast cancer should receive adjuvant chemotherapy has routed through a molecular assay — tissue shipped to a central lab, RNA extracted, a recurrence score returned days later. In May 2026, that workflow gained a new option that never leaves the glass slide. The FDA cleared ArteraAI Breast through the 510(k) pathway, making it the first FDA-cleared digital-pathology AI risk-stratification tool specifically for HR+/HER2− invasive early breast cancer [1].

What's notable here isn't just another algorithm. It's where the input comes from. ArteraAI Breast derives its risk features from hematoxylin and eosin (H&E) whole-slide images — the same stained section a pathologist is already looking at — and generates an image-derived risk score [1]. No separate molecular assay has to be ordered for the tool itself to run. The company positions the score to sit alongside, and potentially complement, existing genomic assays rather than replace them [1]. That framing matters, and we'll come back to it, because the most honest read of this clearance is that it adds a modality to a crowded decision rather than settling it.

The diagnostic problem it addresses

To see why an image-based score is interesting, it helps to remember how much of the HR+/HER2− workup already rides on a handful of markers. The diagnosis itself is built on immunohistochemistry: estrogen receptor (ER) and progesterone receptor (PR) status define the "HR+" half of the label, and HER2 (ERBB2) status — by IHC and, when equivocal, by in situ hybridization — defines the "HER2−" half. The Ki-67 proliferation index often gets folded in as a proxy for tumor aggressiveness. These are the workhorses, and they're reproducible enough to anchor a diagnosis but blunt enough that two tumors with identical ER/PR/HER2/Ki-67 profiles can behave very differently.

That residual uncertainty is exactly the gap multigene assays were built to fill. The 21-gene recurrence score and the 70-gene signature translate tumor biology into a quantitative estimate of recurrence risk and, by extension, expected chemotherapy benefit. They've reshaped practice by sparing many patients chemotherapy they wouldn't benefit from. But they carry real friction: cost, turnaround time, tissue requirements, and uneven access — particularly outside well-resourced centers and health systems.

An H&E-derived score speaks directly to that friction. Because the input is the routine stained slide, the barrier to ordering is lower, and in principle the cost and access profile improves [1]. The clinical substrate it touches is the same one every breast pathologist knows: ER, PR, HER2 (ERBB2), the Ki-67 index, and the genomic recurrence-score framework these AI tools are explicitly benchmarked against [1]. The pitch, in short, is a risk signal extractable from tissue you've already processed.

The evidence

Here's where measured language earns its keep. A 510(k) clearance establishes that the device is substantially equivalent to a predicate and met the FDA's bar for its intended use [1] — it is not, by itself, a claim of superiority over genomic assays, nor a demonstration of long-term outcome benefit in a randomized setting.

The clinical-utility conversation is actively unfolding. At the 2026 ASCO Annual Meeting, Artera presented multiple abstracts on its Multimodal AI (MMAI) approach, which integrates digital pathology with clinical and other data modalities for risk stratification and treatment-decision support across oncology indications, including breast cancer [2]. The data presented were offered to support using image-based AI scores to inform adjuvant decisions — specifically the chemotherapy de-escalation question — and to contextualize where such a score sits relative to genomic assays in the clinical workflow [2]. That last point is the crux of reader interest: not "does it work in isolation," but "what does it add on top of, or in place of, a recurrence score?"

It's worth being precise about the maturity gradient here. The cleared product has a defined, regulated intended use [1]. The MMAI data framing chemotherapy de-escalation and workflow positioning remains investigational [2] — conference abstracts are a signal of direction, not a substitute for peer-reviewed, prospectively validated outcome data. The evidence suggests an image-derived score can carry prognostic information from routine slides; it does not yet tell us, from the records at hand, how that score performs head-to-head against the 21-gene assay across the full spectrum of patients, or how it behaves in the intermediate-risk zone where decisions are hardest.

And one caution that belongs in any digital-pathology discussion: H&E-based models are sensitive to pre-analytic variables — fixation, staining protocols, scanner differences across sites. Generalizability across laboratories is the open question for any image-derived tool, and it's the one to press hardest in local validation.

Where it stands

So where does this leave the practicing pathologist and the multidisciplinary team? With a genuinely new option and a genuinely open set of questions.

The regulatory status is clear enough: ArteraAI Breast is FDA-cleared for risk stratification in HR+/HER2− invasive early breast cancer, the first digital-pathology AI tool to reach that bar in this setting [1]. The MMAI treatment-decision work shown at ASCO 2026 is the research arm pointing toward broader decision support, and it should be read as investigational until the supporting studies mature [2].

The limitations frame the near-term. This is a risk-stratification aid, not a treatment directive — any downstream linkage to adjuvant chemotherapy concerns eligibility for a therapy class and is a decision for the treating oncologist and patient, informed by the full clinical picture, not by a score alone. The relationship to genomic assays is positioned as complementary [1], which means the field now has to work out the practical choreography: Does the image score triage who even needs a genomic assay? Does it resolve cases the markers leave ambiguous? Does it stand alone where genomic testing isn't accessible? None of that is settled.

The honest summary is that a new modality has crossed from research into cleared clinical use, and the hard, interesting work — defining exactly where it sits in the HR+/HER2− pathway, and validating it in your own lab's hands — is just beginning. That's not a hedge. That's the job.

---

References

1. Artera AI Inc. Artera Receives U.S. FDA Clearance for ArteraAI Breast, Expanding Its AI Platform to Breast Cancer. Artera AI Press Release (primary regulatory announcement), 2026. https://artera.ai/news/artera-receives-u-s-fda-clearance-for-arteraai-breast-expanding-its-ai-platform-to-breast-cancer

2. Artera AI Inc. Artera Presents Multiple ASCO 2026 Abstracts on Multimodal AI (MMAI) Risk Stratification and Treatment Decisions in Cancer. Artera AI Press Release / ASCO 2026 Annual Meeting abstracts, 2026. https://artera.ai/news/artera-presents-multiple-asco-2026-abstracts-on-multimodal-ai-mmai-risk-stratification-and-treatment-decisions-in-cancer