An examination lasting a few minutes can sometimes change the entire course of a medical pathway.
When a patient undergoes a screening mammogram, she generally feels nothing. No palpable mass, no particular pain, no symptom worrying enough to have led her to a doctor. This is precisely the principle of screening: looking for an abnormality before it manifests itself.
In breast cancer, this head start matters.
The World Health Organization points out that mortality decreases when breast cancers are detected and treated early. Tumors identified while they are still small and have not spread to nearby lymph nodes are more likely to be treated successfully.
But detecting early does not simply consist of producing an image and spotting an abnormality. Between screening, interpretation, additional examinations, diagnosis and follow-up, an entire pathway is built. And it is within this chain that new decision-support tools are beginning to find their place.
Detecting before symptoms
Breast cancer remains a major public health issue. In 2024, around 2.4 million women were diagnosed with breast cancer worldwide, and nearly 694,000 died from it according to the WHO.
Early detection relies on two complementary approaches.
The first consists of rapidly diagnosing a disease when a symptom appears: a mass, a change in the shape of the breast, a skin change or unusual discharge, for example.
The second is screening in people who do not yet show any apparent sign.
In France, the organized program currently offers a mammogram every two years to women aged 50 to 74 at average risk. Mammograms considered normal or benign at the first reading also undergo a second reading, an important particularity of the French system.
This organization responds to a fundamental difficulty: looking for an abnormality that is sometimes extremely subtle amid a considerable number of normal images.
The challenge is not only to find a cancer
A screening mammogram must constantly strike a balance.
Abnormalities must be detected early enough, without unnecessarily multiplying recalls, additional examinations and biopsies in patients who ultimately do not have cancer.
The performance of screening therefore cannot be reduced to the number of cancers detected alone. Recall rates, false positives, cancers diagnosed between two screening rounds and radiation exposure are also among the criteria used to evaluate imaging strategies. It is notably on this set of parameters that the Haute Autorité de Santé evaluates changes in screening, such as the introduction of breast tomosynthesis.
For the radiologist, this means making a decision in an environment where the majority of examinations are normal, while only a few images may contain a decisive abnormality.
It is precisely in this space that decision-support tools are developing.
A second look at the image
Artificial intelligence applied to mammography does not simply consist of announcing “cancer” or “no cancer.”
Depending on the systems and workflows, it can flag a suspicious region, assign a risk level to an examination or contribute to prioritizing readings. The objective is then to provide the radiologist with additional information at the moment they are building their decision.
Recent results are beginning to go beyond the scope of retrospective studies alone.
The PRAIM study, published in 2025 in Nature Medicine, observed more than 460,000 women in twelve German centers. In this real-world screening context, AI-assisted reading was associated with a detection rate of 6.7 cancers per 1,000 women screened, compared with 5.7 per 1,000 without assistance, with no increase in the recall rate.
In 2026, the results of the large Swedish randomized MASAI trial also provided important data. AI-assisted screening achieved favorable results compared with conventional double reading, notably in terms of sensitivity and cancers diagnosed between two screenings, while substantially reducing the reading workload.
These results do not mean that the algorithm replaces the radiologist.
Rather, they suggest that another form of reading is becoming possible: human expertise enriched by a system capable of drawing attention to certain information in the image.
From detection to the patient pathway
Yet finding an abnormality is only the beginning.
A suspicious image can lead to additional mammographic views, an ultrasound, an MRI or a biopsy. The results must be compared with previous examinations and the clinical context. Then, when a cancer is confirmed, a new succession of decisions begins: characterization of the disease, therapeutic strategy, treatment and surveillance.
Breast cancer thus reveals an often-forgotten dimension of medical imaging: an image gains its value over time.
Comparing a current examination with previous ones makes it possible to look for a change. Measuring a lesion reproducibly helps assess its progression or its response to treatment. Centralizing relevant information can prevent a subtle change from being lost in the succession of examinations.
Decision-support tools can then intervene beyond detection alone. Quantification, longitudinal comparison, structuring of information or highlighting of regions of interest can help the professional more easily follow what changes from one examination to the next.
But the true objective remains the same: maintaining the continuity of the pathway.
The WHO specifically emphasizes the importance of a continuum linking early detection, rapid diagnosis and comprehensive care. It also highlights the value of patient navigation in reducing breaks in the pathway and facilitating timely access to the different stages of diagnosis and treatment.
Helping to decide does not mean deciding
The arrival of AI in screening could give the impression that one day analysis will become entirely automatic.
Clinical reality is more complex.
An abnormality identified on a mammogram must be interpreted according to its morphology, its location, breast density, previous examinations and the patient’s context. An alert produced by a system is therefore not a medical conclusion.
It is additional information.
This distinction is essential to avoid two opposite pitfalls: systematically ignoring the algorithm or, conversely, placing excessive trust in it.
The performance observed in the large recent studies is encouraging, but it depends on the model used, the population, the organization of screening and the way the tool fits into radiologists’ work. Results obtained in one health system cannot be automatically transposed to all others.
The challenge is therefore not to automate the decision. It is to build an environment in which the right information reaches the right professional at the right time.
A few millimeters, sometimes several years ahead
The early detection of breast cancer ultimately tells something very simple.
The more a disease can be identified before it progresses, the more options medicine has to act. The objective of screening is not to produce more images, but to transform a still-silent abnormality into usable information early enough.
Decision-support tools are today extending this logic.
They can draw attention, compare, quantify or contribute to organizing the reading. They can help manage growing volumes of examinations and, in certain contexts, improve screening performance. But their value is not measured solely by that of their algorithm.
It is measured by what happens next.
A detected image must become understood information. Understood information must become, when necessary, an additional examination. Then a diagnosis. Then treatment.
Between the first mammogram and the follow-up of a patient, technology can help shorten certain distances.
But the purpose remains profoundly human: ensuring that when a cancer appears, medicine can see it early enough to have the greatest possible number of options.
