Constance Lehman still remembers the mammogram images of the 75-year-old woman whose AI risk score quietly climbed year after year—2.0 in 2015, 2.1 in 2018, 3.4 in 2019, 3.6 in 2021—before spiking to 15.3 in 2022, the same year a new mass appeared and was confirmed as invasive ductal carcinoma. This woman’s story is no longer an outlier but part of a powerful pattern uncovered by AI: breast cancer risk isn’t static, and its evolution over time can be seen in the very pixels of routine mammograms. In a landmark study published in Radiology, researchers analyzed more than 158,000 mammograms from 54,014 women across urban, community, and rural U.S. imaging centers, revealing that changes in AI-generated risk scores over time are far more telling than a single snapshot ever could. This shift from static to dynamic risk assessment could transform early detection, especially for the 70% of women who develop breast cancer without known genetic risks or family history.
Traditional models have long struggled to predict cancer in the general population, often relying on limited factors like breast density or self-reported family history. But deep learning models, such as the open-source system used in this study, analyze the entire mammogram image, picking up subtle, complex patterns invisible to the human eye. When applied to serial exams, the AI revealed a stark divergence: women who eventually developed cancer saw their median risk score rise from 2.1 to 6.6 in the six years before diagnosis, while scores for cancer-free women remained steady, between 1.8 and 2.2. The difference became statistically significant as early as six years before diagnosis—offering a potential window for earlier intervention.
Among the 817 women diagnosed with cancer, 83% were screen-detected and 17% were interval cancers—those that arise between screenings. The AI’s ability to detect rising risk trajectories even in women with no prior red flags underscores its potential to catch cancers earlier, particularly aggressive or fast-developing ones that current methods might miss. Importantly, the model used no demographic or clinical data—only the images themselves—proving that the mammogram holds more information than previously imagined.
This isn’t just about better algorithms; it’s about redefining how we think about risk. “We observed clinically relevant differences in risk trajectories between women who did and did not develop cancer,” said Dr. Lehman, professor of radiology at Harvard Medical School and CEO of Clairity Inc. As AI tools become more integrated into radiology workflows, the future may allow personalized screening schedules—more frequent checks for those whose scores rise, and less anxiety for those whose remain flat. For millions of women, the quiet climb of a number could one day mean the difference between life and death.