At NYU Langone Health, researchers have trained an artificial intelligence system to read the slow language of changing breast tissue across years of mammograms, offering a glimpse of cancer risk up to five years before diagnosis. Where traditional screening waits for disease to announce itself, this tool listens for quieter signals — patterns in density and structure that accumulate invisibly over time, beyond the reach of family history or genetic testing alone. The work is not yet ready for the clinic, but it points toward a future in which prevention begins long before a tumor forms.
NYU AI tool may predict breast cancer risk up to 5 years early using mammogram patterns
A mammogram can tell us more than whether cancer is visible today
Mark
So the AI is looking at old mammograms and finding patterns that predict cancer years before it shows up. How does it actually do that?
Mimi
It compares a woman's mammograms over time. Breast tissue changes—density shifts, structural changes. The AI learned from hundreds of thousands of scans which patterns tend to precede a cancer diagnosis within five years.
Luke
But we should be clear: the model was trained on data from one hospital system using one manufacturer's equipment. We don't know yet if those patterns hold up elsewhere.
Mark
What's the practical benefit for a patient?
Mimi
If you're flagged as higher risk, you might get an MRI in addition to your regular mammogram. If you're average risk, you stick with routine screening. And the AI can reanalyze scans you've already had—no new radiation.
Luke
That's real. But Lehman was careful to say this isn't a diagnosis and doesn't guarantee anything. A higher score starts a conversation, not a treatment plan.
Mark
Why does this matter more than just looking at family history?
Mimi
Because family history misses people. Lehman said she's treated countless women with no family history who were shocked by a diagnosis. The AI can see risk signals in the tissue itself.
Luke
True, but we're still in research territory. It's not FDA-approved. The next step is testing it on data from other hospitals and different equipment to see if it actually works outside NYU.
Mark
So this could change screening, but we're not there yet.
Mimi
Exactly. It's promising, but it needs validation first.
Luke
And even if it works, it's a tool to inform decisions, not to replace a doctor's judgment or regular screening.
O Pulso
- Millions of women with no family history and no genetic red flags are diagnosed with breast cancer each year, exposing a fundamental gap in how risk is currently understood and communicated.
- NYU Langone's AI model — trained on over 313,000 mammograms from more than 161,000 women — can detect subtle shifts in breast tissue across multiple scans that no single image, and no human eye, would reliably catch.
- The tool requires no new scans or radiation; it reinterprets imaging a woman has already undergone, turning existing medical records into a potential early-warning system.
- A higher AI risk score is not a verdict — it opens a clinical conversation about additional imaging or risk-reduction strategies, while a lower score does not replace the need for routine mammography.
- The model is experimental and not FDA-approved, trained on data from a single hospital system using one manufacturer's equipment — broader validation across diverse centers and imaging platforms is the critical next step.
At NYU Langone Health, researchers have trained an artificial intelligence system to read the slow language of changing breast tissue across years of mammograms, offering a glimpse of cancer risk up to five years before diagnosis. Where traditional screening waits for disease to announce itself, this tool listens for quieter signals — patterns in density and structure that accumulate invisibly over time, beyond the reach of family history or genetic testing alone. The work is not yet ready for the clinic, but it points toward a future in which prevention begins long before a tumor forms.
Researchers at NYU Langone Health have built an artificial intelligence system that reads years of a woman's mammograms in sequence, estimating her likelihood of developing breast cancer within the next five years. Published in the American Journal of Roentgenology, the work marks a meaningful shift in screening philosophy: rather than waiting for cancer to appear, the tool searches for patterns in breast tissue that accumulate gradually — changes a radiologist might never think to connect across separate visits.
The model was trained on more than 313,000 mammograms from over 161,000 women, giving it the scale to recognize subtle variations in tissue density and structure. Its defining innovation is temporal — it tracks how breast tissue evolves across multiple scans, much like a meteorologist reading pressure systems over days rather than consulting a single weather snapshot.
Breast imaging specialist Connie Lehman of Harvard Medical School, who was not involved in the research, underscored why this matters. For decades she has seen women devastated by diagnoses despite having no family history and no known genetic mutations — nothing that traditional risk models would have flagged. An AI trained on hundreds of thousands of images can detect signals embedded in tissue itself, signals that exist entirely outside the domain of genetics or family patterns.
The practical implications are significant. Women identified as higher risk could pursue additional imaging such as MRI; those at average risk could continue routine screening. Crucially, the tool can reanalyze mammograms a woman has already had — no new scans, no additional radiation. Lehman was careful to draw clear boundaries: a risk score is not a diagnosis, and a lower score does not eliminate the need for regular mammography. The tool is a lens, not a crystal ball.
The model remains experimental and lacks FDA authorization. It was trained on data from NYU Langone hospitals using equipment from a single manufacturer — a real limitation. Before clinical deployment, it must be validated across different medical centers, patient populations, and imaging systems. That broader testing is planned, and its outcome will determine whether the patterns learned in one hospital system hold up in the wider world.
Researchers at NYU Langone Health have developed an artificial intelligence system that can look backward through years of a woman's mammograms and estimate her risk of developing breast cancer within the next five years. The work, published in the American Journal of Roentgenology, represents a shift in how screening might work: instead of waiting for cancer to appear on a scan, the tool hunts for patterns in breast tissue that accumulate over time, patterns a radiologist might miss or never think to connect.
The model was built on a substantial foundation. Researchers trained it using more than 313,000 mammograms collected from more than 161,000 women. That scale of data allows the system to recognize subtle shifts in tissue density and structure that correlate with future cancer risk. The key innovation is temporal—the AI doesn't just look at today's image. It watches how a woman's breast tissue has changed across multiple scans, the way a meteorologist might track pressure systems across days rather than taking a single snapshot of the sky.
Connie Lehman, a breast imaging specialist at Harvard Medical School who was not part of the research, explained the practical significance. For decades, she has encountered women devastated by a breast cancer diagnosis who had no family history, no genetic mutations, nothing in the traditional risk profile that would have flagged them for closer attention. "How can this be?" they ask. "No one in my family has had breast cancer." Traditional risk assessment relies on factors like family history and known genetic mutations—important anchors, but incomplete ones. An AI system trained on hundreds of thousands of mammograms can detect additional signals embedded in the tissue itself, signals that exist independent of genetics or family patterns.
The practical implications could reshape screening. Women identified as higher risk might pursue additional imaging, such as MRI, in conversation with their doctors. Women at average risk could continue with routine mammograms. The tool has another advantage: it can reanalyze mammograms a woman has already had. There is no need for new scans, no additional radiation exposure. The information is already there, waiting to be read in a different way.
Lehman emphasized important boundaries. An AI risk score is not a diagnosis. It cannot predict with certainty whether someone will develop cancer. A higher score opens a conversation with a doctor about screening options and risk reduction. A lower score does not eliminate risk or replace regular mammography. The tool is a lens, not a crystal ball.
The model remains experimental. It is not yet authorized by the FDA for clinical use. The researchers trained it on mammograms from NYU Langone hospitals using equipment from a single manufacturer. That limitation matters. Before this tool could move into widespread clinical practice, it needs to be tested on data from other medical centers and with different 3D mammography systems. The team has announced plans to conduct that broader validation. The question now is whether the patterns the model learned from one hospital system will hold up when applied to different populations, different equipment, different imaging protocols. That work is ahead.
Citações Notáveis
For decades, as a doctor and breast imaging specialist, I have cared for women who were shocked by a breast cancer diagnosis. Again and again, I heard: 'How can this be? No one in my family has had breast cancer.'— Connie Lehman, breast imaging specialist, Harvard Medical School
A higher score can prompt a conversation with your health care provider about additional screening and risk-reduction options. A lower score does not eliminate risk or replace regular mammography.— Connie Lehman
Quer a matéria completa?
Leia o original em Fox News ↗
Education