A spatial AI-based biomarker with independent prognostic value in early low- and intermediate- risk breast cancer

Breast cancer is a disease where the body's own cells begin to grow out of control, and for many women, the most common type involves cells that are driven by estrogen but do not react to a protein called HER2. Doctors have long known that some of these cancers are very slow-growing and unlikely to return after surgery, while others are more aggressive. To decide who needs extra treatment, such as chemotherapy, pathologists look at tissue samples under a microscope. They assign a grade to the tumor based on how abnormal the cells look; a low grade usually means a better outlook, while a high grade suggests a higher risk. However, this system is not perfect. Many patients with tumors that look only moderately abnormal still face a risk of the cancer coming back years later, while others with similar-looking tumors remain completely healthy. The challenge is finding a way to tell the difference between these two groups without relying on expensive genetic tests or guessing.

A team of researchers in Sweden has developed a new way to look at these tissue samples using artificial intelligence. Instead of just giving the whole tumor a single average score, their method pays close attention to where specific patterns appear on the slide. They discovered that the location of dangerous-looking cells matters just as much as how many there are. Specifically, they found that when clusters of high-risk cells gather at the very edge of the tumor, where it meets the healthy tissue, the risk of the cancer returning is significantly higher. This finding offers a new tool to identify which patients truly need more aggressive treatment and which ones can safely avoid it, all by using the standard microscope slides that hospitals already have.

The researchers started with a simple idea: that the way a tumor is organized in space holds secrets that a simple average score misses. Imagine a tumor not as a uniform blob, but as a landscape with different neighborhoods. Some parts might look calm and orderly, while others are chaotic and aggressive. For years, artificial intelligence models used to analyze these slides have been trained to look at the entire image and produce one single number representing the overall risk. While these models are powerful, they often smooth over the details, blending the chaotic edges with the calm centers into a single, diluted score. The team behind this study, working with data from thousands of patients across Sweden, hypothesized that if they could map exactly where the dangerous parts were located, they might find a clearer signal. They focused their attention on the "invasive front," the outer boundary of the tumor where cancer cells interact with and push into the surrounding healthy tissue. This is a biologically active zone where cells often change shape to become more mobile and invasive.

To test this, the scientists used a massive dataset of tissue samples from over 6,000 women with early-stage breast cancer. They divided these patients into two groups: one group to train their computer models and another completely separate group to test the results, ensuring the findings were not just a lucky guess. They used a sophisticated AI system to break each microscope slide into thousands of tiny squares, or tiles, and analyzed the shape and structure of the cells in each one. The AI was trained to recognize patterns that look like high-grade, aggressive cancer. Once the AI had scored every single tile, the researchers did something different from previous studies. Instead of just counting how many high-risk tiles existed, they looked for clusters. They asked a specific question: are there groups of these high-risk tiles sitting right next to each other at the very edge of the tumor? They defined a "cluster" as a group where at least two-thirds of the neighboring tiles were also high-risk. If such a cluster existed at the tumor's front, the patient was marked as "cluster-positive."

The results were striking. In the group of patients with tumors that were generally considered low-to-intermediate risk, the presence of these clusters at the tumor edge completely changed the outlook. Among patients who had these clusters, the risk of the cancer returning within ten years was significantly higher. In fact, the data showed that patients with these clusters had a ten-year survival rate without recurrence of about 90 percent, which is lower than the 92.6 percent rate observed in the Grade 3 reference group. On the other hand, patients who did not have these clusters at the tumor edge had an excellent outlook, with a ten-year survival rate of over 95 percent. This group of "cluster-negative" patients fared just as well as those with the very lowest-grade tumors, suggesting they could be treated with less aggressive therapy.

Crucially, this new marker provided information that the standard AI score could not see on its own. The researchers found that even among patients whose tumors were classified as low-risk by the overall AI score, those who had these specific clusters at the edge still faced a much higher risk of recurrence. This means the spatial arrangement of the cells was revealing a hidden danger that a simple average score had missed. The study confirmed these findings in an independent group of patients from a different region of Sweden, proving that the result was not specific to one hospital or one set of slides. The presence of these clusters was linked to a nearly three-fold increase in the risk of recurrence, a strong signal that held true even when the researchers accounted for other factors like the patient's age or the size of the tumor.

This work suggests that the future of cancer risk assessment may lie in looking at the map of a tumor, not just its average score. By preserving the spatial information of where high-risk cells are located, doctors may be able to spot the patients who are truly at risk, even if their tumor looks mostly harmless. For the patients who do not have these clusters, the news is equally important: it offers strong evidence that they are in a very low-risk group, potentially sparing them from unnecessary treatments. The study does not claim to have solved the problem of breast cancer, but it provides a clear, concrete method to refine how doctors understand risk. It shows that in the complex landscape of a tumor, the location of the danger is just as important as the danger itself.