AI Improves Accuracy in Identifying HER2-Low, Ultralow Breast Cancers
A new international study presented at the 2025 American Society of Clinical Oncology Annual Meeting found that artificial intelligence tools can significantly improve the accuracy of pathologists in identifying HER2-low and HER2-ultralow breast cancers, subtle but clinically meaningful subtypes that could open the door to more targeted treatment options.
The most clinically relevant science in oncology will be presented at #ASCO25, and you can now view the full 5,000 abstracts & their presentation date, time & location info: https://t.co/M6DJtJSwU2 pic.twitter.com/vo5dTbv5Qt
— ASCO (@ASCO) May 22, 2025
The findings have important implications for expanding access to trastuzumab deruxtecan (Enhertu), a HER2-directed antibody-drug conjugate that has shown benefit in patients who were historically considered HER2-negative. These results could help ensure that more patients are accurately identified and considered for therapies that may improve survival and quality of life.
HER2, or human epidermal growth factor receptor 2, is a protein found on the surface of breast cancer cells that helps them grow. Traditionally, breast cancers have been classified as either HER2-positive or HER2-negative based on the amount of this protein. However, recent research has revealed that patients whose tumors express low or ultralow levels of HER2 — not enough to meet the threshold for HER2-positive disease — may still respond to targeted therapies.
Accurately classifying tumors in the HER2-low and HER2-ultralow range is difficult. The visual differences under a microscope are often subtle, and assessments can vary between pathologists. Misclassification can prevent patients from receiving therapies that may benefit them.
To investigate whether AI could reduce this diagnostic uncertainty, researchers developed a digital pathology training platform known as ComPath Academy. The study enrolled 105 pathologists from 10 countries, each of whom reviewed 20 digital images of breast cancer cases across three exams. During the third exam, participants were provided with AI guidance to assist in interpreting HER2 immunohistochemistry results.
With AI assistance, the accuracy of identifying HER2 clinical categories increased from 90.1% to 95.0%. Concordance among pathologists — a measure of consistency — rose from 0.494 to 0.732. Importantly, the number of HER2-low or HER2-ultralow tumors incorrectly classified as HER2-null decreased by 24.4%, potentially reducing the number of patients who would be incorrectly deemed ineligible for HER2-directed therapies.
Marilyn Bui, MD, PhD
“Pathology and laboratory medicine serves as the cornerstone of precision medicine, delivering accurate and timely diagnostic information and identifying key biomarkers that inform clinical decision-making..” said Marilyn Bui, MD, PhD, scientific director of the Analytic Microscopy Core at Moffitt Cancer Center. “As pathologists, we embrace technologies that enhance diagnostic accuracy and improve our workflow.”.”
The researchers say the use of AI could be particularly valuable in regions where resources or specialist experience are limited. By guiding pathologists toward more consistent and accurate interpretation of HER2, the technology may help level the playing field in breast cancer care.
AI continues to gain traction across many areas of oncology, from screening and diagnostics to treatment planning and clinical decision support. This study adds to the growing body of evidence showing the promising potential of integrating digital tools into pathology workflows to support personalized treatment strategies.
