Why help radiologists read mammograms?
In Germany, every screening mammogram is read by two radiologists. Most images are normal, which makes the work very repetitive, and screening programs are short on radiologists. The goal is to miss as few cancers as possible while avoiding false alarms, which cause needless worry.
Studies on past images had suggested AI could help. But it still had to be tested in everyday practice, with real radiologists, real machines, and real patients.
What did the researchers do?
From July 2021 to February 2023, 12 German screening sites took part, with 119 radiologists and equipment from five manufacturers. For each exam, the radiologist chose whether to read it with or without the AI software Vara MG, a CE-marked medical device in Europe. In total, 260,739 women were read with AI and about 201,000 without.
The AI had two jobs. It tagged as "normal" the exams it found very reassuring. It also acted as a "safety net": if a radiologist called an exam normal but the AI found it highly suspicious, an alert asked them to take another look at the flagged area. The final decision was always made by a person.
The results: more cancers found, no more callbacks
With AI, radiologists detected 6.7 cancers per 1,000 women, vs 5.7 without AI: one more cancer per 1,000 women, or 17.6% more. The recall rate, meaning women called back for more tests, stayed stable at 37.4 per 1,000 with AI and 38.3 without.
Among women called back, more actually had cancer: 17.9% with AI vs 14.9% without. The safety net went off 3,959 times and was accepted 1,077 times, leading to 204 cancer diagnoses that the authors say would otherwise have been missed. Radiologists also read exams tagged "normal" faster.
Key figures
- 463,094women screened at 12 German sites, 260,739 of them with AI support
- +17.6%more breast cancers detected with AI (6.7 vs 5.7 per 1,000 women)
- 37.4 per 1,000women recalled with AI, vs 38.3 without: no increase in recalls
- 204breast cancers diagnosed after an AI "safety net" alert
Your breast screening, step by step with Mima
In Germany, AI already helps radiologists read screening mammograms, without replacing their decision. Depending on where you live and your imaging center, this kind of help may or may not be in use yet.
Mima helps you log your mammograms, their results, and any follow-up tests, so you can keep track from one year to the next. She also prepares your questions, such as: "If software flags an area on my images, who reviews it, and how?"
What this means for patients
This study suggests that AI, used as a helper rather than a replacement, can help find more breast cancers without adding stressful callbacks. The authors call for this kind of tool to be built into screening guidelines.
If you are due for a screening mammogram, a simple question to ask is: does my imaging center use AI to help read the images?
What are the limits of this study?
This was not a randomized trial: radiologists chose whether to use AI, which created a bias the researchers corrected with statistical methods. The study was funded by Vara, and three authors work for the company.
AI mostly led to finding more ductal carcinoma in situ (very early cancers that stay inside the milk ducts), which raises the question of overdiagnosis. It is not yet known whether this screening will reduce cancers found between mammograms or deaths: a 2 to 3 year follow-up is planned.
Frequently asked questions
Is AI used in mammogram screening?
Yes, in some programs. In Germany, the PRAIM study used it for more than 260,000 women to support two radiologists, and more cancers were detected.
Does AI replace the radiologist in mammography?
Not in this study. The AI flagged normal or suspicious exams, but the decision to call a woman back was always made by radiologists.
Does AI increase false positives on mammograms?
Not in PRAIM: the callback rate was 37.4 per 1,000 women with AI vs 38.3 without, and a larger share of women called back really had cancer.
What is ductal carcinoma in situ (DCIS)?
It is a very early breast cancer that stays inside the milk ducts. AI led to finding more of them, which raises the question of overdiagnosis.
The original study
Eisemann N, Bunk S, Mukama T, et al. “Nationwide real-world implementation of AI for cancer detection in population-based mammography screening” Nature Medicine, January 7, 2025. DOI : 10.1038/s41591-024-03408-6
Read the full study in Nature MedicineThis article summarizes a scientific study and is not medical advice.