Mammograms were already snitching on future breast cancer, and when researchers added DNA receipts, the predictions got better.

That is the headline from a 2026 study in the British Journal of Cancer, and yes, it has strong "the walls have ears" energy. Researchers tested whether an image-only deep learning model called Mirai could predict who would develop breast cancer within five years, then asked an obvious follow-up: what if we also toss in a polygenic risk score, meaning a genetic risk estimate built from lots of tiny DNA variants that each contribute a little risk? Turns out the combo did better than the image model alone, and much better than the old-school Gail model even after Gail got the same genetic upgrade [1].

A mammogram is not just a yes-or-no cancer detector. It is also a giant pile of visual information: tissue patterns, density, subtle textures, and other details that your eyeballs and my eyeballs would not summarize into a neat risk score before last call. Deep learning models like Mirai are trained to find those patterns and turn them into a forecast of future cancer risk, not just whether cancer is visible right now [2,3].

In this study, the researchers looked at 902 women from the Nurses' Health Study II, including 270 who developed breast cancer and 632 controls. Mirai alone reached an AUC of 0.66. When they added a polygenic risk score based on 313 breast-cancer-associated genetic variants, the AUC rose to 0.73. That is not "we solved cancer with vibes and matrix multiplication." It is a modest but meaningful bump. Meanwhile, the Gail model went from 0.52, which is basically "shrug with paperwork," to 0.69 after adding the same polygenic score. Mirai plus genetics still won [1].

Interesting, right? The image already knew something. The genes knew something else. Put them together and suddenly the risk estimate gets less confused. Coincidence? I am legally required to say yes, statistically modeled coincidence. Spiritually, though, this is a classic "two informants, one conspiracy board" situation.

Polygenic Risk Scores: DNA, but in Spreadsheet Form

A polygenic risk score, or PRS, is what happens when genetics stops looking for one dramatic villain gene and instead rounds up hundreds of tiny accomplices. Each single-nucleotide polymorphism, or SNP, usually has a very small effect. Bundle enough of them together, and you get a more useful estimate of inherited risk [4].

That matters because breast cancer risk is messy. Family history helps. Age helps. Breast density helps. But none of those gives you the whole picture. PRS has been gaining traction because it can sharpen risk estimates and potentially help decide who might benefit from earlier screening, more frequent screening, or preventive interventions [4,5].

What this new paper suggests is that the future may belong to hybrid models. Not "just genes." Not "just images." Not "just a questionnaire you filled out while hungry in a waiting room." More like all of the above, stitched together into something less blunt.

Follow the Screening Strategy

This is where the real-world angle shows up wearing sunglasses indoors. Breast cancer screening guidelines already moved in the US. The USPSTF updated its recommendation on April 30, 2024, to advise biennial mammography for average-risk women ages 40 to 74 [6]. But "average-risk" is doing a lot of emotional labor there. Some women are not average. Some are higher risk and may need earlier or more tailored screening.

That is exactly why AI risk models are getting so much attention. A 2023 Radiology study found several mammography AI models outperformed a standard clinical model for 5-year breast cancer risk prediction [3]. A 2025 systematic review in JNCI looked across 41 studies and found mammography-based AI models typically landed around a median AUC of 0.71, while also warning that calibration and diversity remain major issues [2]. Translation: the models are promising, but some may still be better at ranking risk than telling you the true absolute risk. That distinction is not academic nitpicking. It is the difference between "who is likely higher risk" and "who crosses a clinical threshold that changes care."

Meanwhile, this is no longer living entirely in journal-land. RSNA highlighted 2025 data showing AI risk models beat breast density alone for five-year risk stratification, and by February 2026 Beth Israel Deaconess had already rolled out an AI-driven breast cancer risk tool in practice [7,8]. Follow the money? Fine. Also follow the workflow.

The Fine Print Nobody Should Ignore

Before anybody starts declaring the mammogram a crystal ball with FDA paperwork, slow down.

This study was relatively small, nested case-control, and drawn from a specific cohort [1]. PRS performance can vary by ancestry because many genetic scores were built mostly in European populations, which is a polite way of saying medicine has another diversity homework assignment overdue since forever [4,5]. AI image models also need stronger prospective validation, better calibration, and testing across different machines and patient populations [2]. Helpful? Yes. Ready to run your life by itself? Absolutely not.

Still, the core idea is hard to ignore: your breast imaging and your genetics may each carry different hints about future cancer, and combining them could make screening less one-size-fits-all. For a field that has spent decades balancing early detection, false positives, cost, and access, that is a pretty big deal. Not a magic trick. More like finally getting two blurry security cameras pointed at the same alley.

References

1. Azam S, Lamb LR, Eliassen AH, et al. Performance of an image-only deep learning breast cancer risk model with the addition of a polygenic risk score. British Journal of Cancer. 2026. DOI: 10.1038/s41416-026-03415-z. PubMed: PMID 41942608

2. Lee HEJ, Lowry KP, Marinovich ML, et al. Current state of mammography-based artificial intelligence for future breast cancer risk prediction: a systematic review. JNCI: Journal of the National Cancer Institute. 2025. DOI: 10.1093/jnci/djag002

3. Arasu VA, Habel LA, Achacoso NS, et al. Comparison of Mammography AI Algorithms with a Clinical Risk Model for 5-year Breast Cancer Risk Prediction: An Observational Study. Radiology. 2023;307(5):e222733. DOI: 10.1148/radiol.222733. PubMed: 37278627

4. Roberts E, Howell S, Evans DG. Polygenic risk scores and breast cancer risk prediction. The Breast. 2023;67:71-77. DOI: 10.1016/j.breast.2023.01.003. PMCID: PMC9982311

5. Hughes E, Tshiaba P, Gallagher S, et al. Development and Validation of a Breast Cancer Polygenic Risk Score on the Basis of Genetic Ancestry Composition. JCO Precision Oncology. 2022. DOI: 10.1200/PO.22.00084. PubMed: 36331239

6. U.S. Preventive Services Task Force. Recommendation: Breast Cancer: Screening. April 30, 2024. USPSTF

7. RSNA. AI Tops Density in Predicting Breast Cancer Risk. 2025. RSNA news release

8. Axios Boston. Beth Israel starts using AI-driven tool to spot breast cancer risk. February 23, 2026. Axios

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.