AI Reads Routine Mammograms to Predict Breast Cancer HER2 Status Without Biopsy

Every year, millions of women undergo mammography as part of routine breast cancer screening, and the images generated in those few seconds contain far more information than radiologists can consciously perceive. A new dual-center study from China suggests that a deep learning model trained on those very images can predict, noninvasively, whether a breast tumor carries the HER2 protein — and, remarkably, can even distinguish subtle intermediate levels of the receptor that have only recently become therapeutically important. The research, published in BMC Medical Imaging, hints at a future in which a standard screening mammogram could help guide treatment decisions before a single needle touches the patient.

HER2, or human epidermal growth factor receptor 2, is one of the most consequential molecules in breast oncology. In roughly 15 to 20 percent of breast cancers, the gene encoding this receptor is amplified, driving aggressive tumor growth but also rendering the cancer vulnerable to highly effective targeted therapies such as trastuzumab. Determining HER2 status has traditionally required biopsy followed by immunohistochemistry or fluorescence in situ hybridization, laboratory techniques that are invasive, time-consuming, and subject to sampling error. Because tumors are biologically heterogeneous, a core needle sample may not represent the full molecular landscape of the disease, and repeated biopsies add cost, discomfort, and risk.

The research team, led by Huizhe Wang, Changjun Liu, Yan Zhao, and colleagues at Qingdao Central Hospital and the Affiliated Hospital of Qingdao University, set out to bypass the biopsy altogether. Their strategy belongs to a rapidly growing field called radiomics, which converts medical images into high-dimensional quantitative data. Unlike a human reader who might describe a mass as spiculated or dense, a radiomics pipeline extracts hundreds of mathematical features — texture patterns, intensity distributions, shape descriptors — that correlate with underlying tumor biology. When combined with deep learning, which can learn its own features directly from pixel data, the approach becomes a powerful microscope into tissue characteristics invisible to the eye.

Technically, the study is notable for its choice of architecture. The team employed Visformer, a hybrid vision transformer that blends convolutional operations with self-attention mechanisms, allowing the model to capture both local image details and long-range spatial relationships within the mammogram. Full-field digital mammography images in craniocaudal and mediolateral oblique views were processed, and regions of interest corresponding to the tumor were analyzed. To make the final prediction more robust, the researchers used a weighted voting ensemble, in which multiple model outputs were combined with weights optimized exclusively within the training cohort through cross-validated area under the receiver operating characteristic curve, a design intended to prevent information leakage and overfitting.

The dataset comprised 546 breast cancer patients drawn from two hospitals. Center 1 contributed 420 patients, split into a training cohort of 294 and an internal validation cohort of 126, while Center 2 provided an independent external validation cohort of 126 patients. This dual-center design matters enormously in machine learning research: models that perform well only on the data they were trained on are clinically useless, so testing on patients from a different institution, with potentially different scanners and imaging protocols, is the gold standard for demonstrating generalizability.

The researchers framed two distinct classification problems. The first, and more conventional, task was to distinguish HER2-positive from HER2-negative tumors. Here the model achieved an area under the curve of 0.836 in the training set, 0.763 in internal validation, and 0.714 in external validation. An AUC of 0.5 represents random guessing, while 1.0 represents perfect discrimination, so a value above 0.7 on unseen external data indicates genuinely useful, if imperfect, predictive power. For context, many published imaging biomarkers fail to clear this bar when moved outside their home institution, making the external result the study’s most important number.

The second task is where the study breaks new ground. Recent clinical trials have established that patients whose tumors express low levels of HER2 — a state once classified simply as negative — can benefit from a new generation of antibody-drug conjugates such as trastuzumab deruxtecan. This has created an urgent need to identify HER2-low disease preoperatively, yet the category sits in a diagnostic gray zone that pathologists find challenging. The deep learning radiomics model distinguished HER2-low from HER2-zero tumors with AUCs of 0.890 in training, 0.785 in internal validation, and 0.713 in external validation. In other words, subtle signatures of receptor expression are genuinely encoded in the mammographic appearance of the tumor, and an algorithm can recover them.

Beyond the imaging model, the team performed clinicopathological analyses that revealed which conventional factors independently predict HER2 status. For HER2-positive disease, histological grade carried an odds ratio of 2.28, the Ki-67 proliferation index an odds ratio of 4.71, and BI-RADS category an odds ratio of 2.49. For HER2-low status, progesterone receptor positivity carried an odds ratio of 4.48, BI-RADS category an odds ratio of 1.75, and breast density an inverse odds ratio of 0.32, meaning denser breasts were less likely to harbor HER2-low tumors. All of these associations reached statistical significance. These findings align with known biology: HER2-driven tumors tend to divide rapidly and grade poorly, while the HER2-low category shows hormonal relationships that could help clinicians prioritize patients for advanced testing.

The clinical implications are considerable. If validated prospectively, a mammography-based HER2 predictor could serve as a triage tool at the moment of diagnosis. A patient whose screening images suggest HER2-positive disease might be fast-tracked for confirmatory testing and early referral for targeted therapy planning, while the identification of probable HER2-low status could flag candidates for antibody-drug conjugate strategies down the line. Because mammography is already performed routinely, widely accessible, and inexpensive relative to advanced imaging such as MRI, the marginal cost of adding such an algorithm to existing workflows would be minimal — a crucial consideration for health systems in low- and middle-income countries where molecular diagnostics remain scarce.

Caution is still warranted. The study was retrospective, and the authors themselves emphasize that prospective validation is needed before clinical deployment. AUCs in the low 0.7 range on external data, while respectable, fall short of the reliability required for autonomous decision-making, and the model would function best as an adjunct alongside pathology rather than a replacement for it. Questions about scanner variability, image quality, and performance across diverse populations remain open. Still, the work adds to a compelling body of evidence that routine medical images encode molecular information once thought accessible only through the laboratory. As deep learning methods mature and datasets grow, the humble mammogram — already a lifesaving screening tool — may quietly become one of oncology’s richest sources of molecular intelligence, delivering personalized cancer care at the speed of light.

Subject of Research: Mammography-based deep learning radiomics for noninvasive prediction of HER2 expression, including HER2-low status, in breast cancer

Article Title: Mammography-based deep learning radiomics for predicting HER2 expression in breast cancer: a dual-center study with focus on HER2-low status

Article References: Mammography-based deep learning radiomics for predicting HER2 expression in breast cancer: a dual-center study with focus on HER2-low status. (n.d.). https://doi.org/10.1186/s12880-026-02921-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02921-z

Keywords: breast cancer, HER2, HER2-low, mammography, deep learning, radiomics, Visformer, machine learning, biomarkers, cancer imaging, predictive medicine, antibody-drug conjugates