Mammography-Based Deep Learning Radiomics for Predicting HER2 Expression in Breast Cancer: A Dual-Center Study with Focus on HER2-Low Status
This dual-center study demonstrates that a mammography-based deep learning radiomics model using the Visformer architecture can noninvasively and reliably predict both HER2-positive versus HER2-negative status and HER2-low versus HER2-zero status in breast cancer patients.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Breast cancer remains the most common malignancy affecting women globally, a disease where the specific biological makeup of a tumor dictates the path forward for treatment. Among the many markers doctors examine to understand a tumor, one called HER2 is particularly critical. In roughly one out of every five newly diagnosed cases, the cancer cells carry an abundance of this protein, a condition known as HER2-positive. These tumors tend to grow more aggressively, but they also respond well to specific targeted therapies designed to attack that protein. For decades, the standard approach to identifying these tumors has been to take a small tissue sample, known as a biopsy, and examine it under a microscope. However, a newer category has recently emerged in medical understanding: HER2-low. These tumors do not have enough of the protein to be called positive, yet they possess just enough to potentially respond to a new generation of powerful drugs. Distinguishing between tumors that have no HER2 at all and those that have a little bit is now a vital clinical question, but the traditional method of testing relies on invasive procedures that can sometimes miss the mark due to the uneven distribution of cells within a tumor.
A team of researchers set out to see if a different, non-invasive approach could solve this problem. They turned to mammography, the standard X-ray imaging used for breast cancer screening, which is widely available and familiar to most women. Instead of relying solely on a radiologist's visual inspection of the image, the team applied a technique called deep learning radiomics. This process involves feeding the digital mammogram images into a computer system capable of detecting subtle patterns and textures that are too fine for the human eye to perceive. The researchers wanted to know if these hidden digital signatures could predict the HER2 status of a tumor before any surgery or biopsy took place. They focused on two specific challenges: first, distinguishing tumors that are HER2-positive from those that are not, and second, separating the newly important HER2-low group from those with zero HER2 expression.
To test their idea, the researchers gathered data from 546 women with confirmed breast cancer across two different medical centers. They divided the patients into groups to train their computer models and then to test them, ensuring the results were not just a lucky guess on a single set of data. The computer system analyzed the mammograms, extracting thousands of tiny numerical details about the shape, texture, and density of the tumors. By combining the predictions of several different mathematical models, the team created a single, robust tool designed to make the final call. The results showed that this digital approach could indeed "see" the difference. When trying to tell HER2-positive tumors apart from the rest, the model performed with a high degree of accuracy in its initial tests, and while its performance dipped slightly when tested on a completely new group of patients from a different hospital, it still maintained a reliable ability to distinguish the groups.
The findings were even more encouraging when the team focused on the difficult task of separating HER2-low tumors from those with no HER2 at all. In the initial training phase, the model was highly successful, and it held its ground reasonably well when applied to the external group of patients. The study also looked at the traditional factors doctors use to assess risk, such as the size of the tumor, the patient's age, and the density of the breast tissue. They confirmed that certain characteristics, like a high grade of tumor cells and a rapid rate of cell division, were strongly linked to HER2-positive status. Similarly, they found that hormone receptor status and breast density played a role in identifying HER2-low tumors. However, the computer model's ability to predict these statuses directly from the X-ray images offered a new layer of insight that did not depend on these other variables alone.
The researchers acknowledged that their work has boundaries. Because the study looked back at past patient records rather than following new patients forward in time, there is a chance that the selection of data introduced some bias. The manual outlining of the tumor areas on the X-rays by human radiologists also means the process is not yet fully automated. Furthermore, the number of patients, while substantial, was not large enough to guarantee the results would hold true for every possible scenario. Despite these limitations, the study suggests that mammography, often viewed simply as a tool for finding cancer, might also serve as a powerful window into the molecular nature of the disease. By offering a non-invasive way to estimate HER2 status, this approach could one day help doctors make more informed decisions about treatment plans, potentially reducing the need for repeated biopsies and ensuring that patients receive the right targeted therapies sooner. The work does not claim to replace the current gold standard of tissue testing, but it points toward a future where imaging and artificial intelligence work together to provide a clearer, more complete picture of the disease before a single needle is ever used.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.