AI Assistance May Boost Accuracy of HER2-Low Scoring for Breast Cancers

CHICAGO — Artificial intelligence (AI) assistance appeared to enhance pathologists’ accuracy in identifying HER2-low or -ultralow expression in breast cancer tissue, a global study showed, which could cut down on missed opportunities for effective HER2-targeted therapy.

Institution of AI as part of a training program to assist pathologists in HER2 scoring of breast cancer samples was associated with a relative 24.4% fewer HER2-low or -ultralow cases being misclassified as HER2-null (complete absence of HER2 expression), reported Marina De Brot, MD, PhD, of A.C. Camargo Cancer Center in São Paolo, Brazil.

Only 3.9% of immunohistochemistry (IHC) readings from whole slide images were inaccurate with AI assistance, whereas 10.9% of readings were inaccurate without it, according to findings presented at a virtual press briefing ahead of next week’s American Society of Clinical Oncology (ASCO) annual meeting.

Improved accuracy in HER2 scoring has important implications for treatment, said De Brot. The reduction in misclassifications of HER2-low and HER2-ultralow cases as HER2-null potentially enables more patients to have access to effective HER2-directed antibody-drug conjugate (ADC) therapy, she said.

AI can potentially reduce pathologist workload while also enhancing patient care, said Julie R. Gralow, MD, ASCO chief medical officer and press briefing moderator.

“With the increasing use and application of anti-HER2-directed therapies, especially in the [HER2]-low and -ultralow populations, patients now have access to potentially life-changing medications” by accurate HER2 classification, she said.

In the DESTINY-Breast06 trial, the ADC trastuzumab deruxtecan (Enhertu) significantly improved progression-free survival in previously treated metastatic hormone receptor-positive/HER2-low or -ultralow breast cancer to a median of 13 months compared with 8 months with chemotherapy alone, leading to expanded approval for that population.

The discordance rate amongst pathologists in evaluating HER2-low and HER2-ultralow breast cancers can reach 30%. “Our results show that across almost 2,000 readings, pathologists achieved a higher accuracy compared to the reference IHC score with AI support versus without AI support,” she said. “We also saw an increase in concordance amongst pathologists with AI support.”

The study compared HER2 scoring of 20 digital IHC-stained breast cancer samples with and without the use of AI assistance against “ground truth” IHC scores — considered to be the gold standard — as defined by a central reference center of expert breast pathologists from multiple institutions. All pathologists applied ASCO/College of American Pathologists 2023 guidelines adapted to include HER2-low and -ultralow breast cancer definitions. HER2-low is defined as an IHC score of 1+ or 2+ with a negative fluorescence in situ hybridization result. HER2-ultralow is characterized by an IHC score of 0 with more than 0% but no more than 10% staining in tumor cells.

A total of 105 pathologists from 10 countries in Asia and South America with varying degrees of experience in HER2 assessment participated. Over the course of five sessions, the pathologists performed 1,940 readings that were done during three separate exams. After the first exam with manual reading, they had a lecture on scoring, followed by another manual test, then discussion of the results of the first two tests. Afterward, they took a third test with AI support.

With AI support, HER2-low scoring sensitivity increased from about 78% to 90%, and pathologists’ agreement with central reference scores improved by about 13%. The average agreement with the central reference scores also improved to 89.6% with AI assistance, compared to 76.3% without AI assistance.

Accuracy in correctly identifying cases as HER2-positive, HER2-low, HER2-ultralow, or HER2-null improved from 90.1% without AI to 95% with AI. AI support raised sensitivity across HER2 clinical classifications, from 54.08% to 88.24% in the case of HER2-null and from 50.74% to 93.22% for HER2-ultralow. HER2-ultralow underscoring occurred in 30.5% of instances with manual scoring but in only 4.5% with AI.

Concordance among pathologists for HER2 clinical categories improved from 0.494 without AI to 0.732 with AI, whereas concordance for HER2 IHC scoring improved from 0.506 without AI to 0.798 with AI assistance.

Improvements in scoring and classification accuracy were realized regardless of pathologist experience in HER2 assessment, said De Brot.

One limitation of the study was potential for confounding of the results by learning, as AI was only offered on the final exam whereas manual reading scores pooled together the results from before teaching sessions.

Disclosures

The study was funded by AstraZeneca.

De Brot disclosed no relevant relationships with industry. Coinvestigators disclosed relationships with AstraZeneca, Mindpeak, and other entities.

Gralow disclosed no relevant relationships with industry.

Primary Source

American Society of Clinical Oncology

Source Reference: De Brot M “Use of artificial intelligence-assistance software for HER2-low and HER2-ultralow IHC interpretation training to improve diagnostic accuracy of pathologists and expand patients’ eligibility for HER2-targeted treatment” ASCO 2025; Abstract 1014.