For the first time, a study has proven, through a controlled clinical trial, that a system based on artificial intelligence can significantly reduce the appearance of aggressive breast cancers in the interval between screening exams.
The Masai study (mammographic screening with artificial intelligence), published this Thursday (29) in the scientific journal The Lancet, followed around 105 thousand Swedish women and demonstrated that the use of AI in reading mammograms resulted in 12% fewer interval tumors, those diagnosed between one appointment and another — precisely the most dangerous and difficult to treat cases.
Coordinated by physician and researcher Kristina Lång, from Lund University in Sweden, the clinical trial compared two screening methods: on the one hand, the standard European practice, in which two radiologists analyze each mammogram performed; on the other, an artificial intelligence system that first evaluates the exam, forwards low-risk cases to just one specialist and maintains double reading only for suspicious situations. The machine also works as a kind of “second look”, highlighting areas that deserve special attention.
The results were surprising due to their consistency. Not only did the AI detect 29% more tumors than the traditional method, it did so without increasing false positives — which lead healthy women to have biopsies. And it also reduced radiologists’ workload by half.
The system used in the study is not a generic artificial intelligence, it has a name and owner: Transpara, from the Dutch ScreenPoint Medical, already approved by regulatory agencies in Europe and the United States. Trained with more than 200,000 mammography exams from different countries, the software analyzes the images and assigns a malignancy risk score on a scale of 1 to 10, with cases with a score of 10 being sent for reading by radiologists.
So-called “interval cancers” represent one of the greatest weaknesses of screening programs. Estimates indicate that between 20% and 30% of tumors diagnosed after a normal mammogram could have been detected in the previous exam. These cancers tend to grow quickly and are often in more advanced stages when they are finally discovered.
In the study, the group that used AI had 82 interval cancer cases versus 93 in the control group. Most importantly, there were 16% fewer invasive cancers, 21% fewer large tumors (over 2 cm), and 27% fewer cases of the most aggressive subtypes. “Our study is the first randomized controlled trial investigating the use of AI in breast cancer screening and the largest to date on the use of AI in cancer screening in general,” says Kristina Lång.
For the researcher, “widely implementing AI-assisted mammography in breast cancer screening programs could help reduce workload pressures among radiologists, as well as help detect more early-stage cancers, including those with aggressive subtypes.”
Oncologist Laura Testa, from Oncologia D’Or, explains the importance of focusing on these specific tumors. “Interval tumors have the characteristic of being more aggressive, in the sense that they grow more quickly, so there is time for them to show symptoms between one exam and another”, he says.
For Testa, the choice of this outcome makes the study even more clinically relevant. “The fact that we have as the objective of the study this reduction of interval cancer, from a clinical point of view, is very relevant. Because this type of tumor, in fact, is more implicated in a worse prognosis when we let it pass for a long time.”
Testa also raises an important regulatory issue: just as similar medicines need to undergo new clinical studies for validation, different AI systems would also need to demonstrate their own performances. There is no guarantee that all artificial intelligence technologies will perform the same as the system tested in the Swedish study, which means each tool would need to go through validation processes similar to those that generic medicines face.
Oncologist Fernando Maluf, from Einstein Hospital Israelita, agrees with the assessment. “The randomized study, in my opinion, is one of the most relevant on AI in terms of radiology for cancer screening, published so far”, he states. “It puts AI in this scenario as a potential standard of care in reading mammograms.”
If the clinical results are impressive, the economic viability is still unknown. Testa ponders the challenges: “It’s a concern because, just as when we have a new medicine that comes with the promise of an important improvement for patients, this usually comes at a very high cost.”
The main expense would be the AI system itself, but there are potential savings, such as in radiologists’ dedicated hours. A Norwegian modeling study suggested that AI would be cost-effective if it reduced interval cancers by at least 5%, and in the context of the study it was 12%.
Kristina Lång says that “although cost-effectiveness has not yet been assessed, this reduction suggests that the approach has the potential to be cost-effective.” The cost-benefit analysis of the study itself is underway and should be published soon, according to the researcher. She points out that any increase in false-positive recalls would also increase healthcare costs. “However, earlier detection of clinically relevant cancers has the potential to reduce subsequent treatment costs, which could represent a substantial benefit,” he says.
In any case, even though the advances are solid, there are challenges, especially for Brazil. “For me, the big challenge for us to repeat this kind of thing here in Brazil is that we have a very low rate of mammographic screening. Mammographic screening is not how many mammograms we do per year. It’s how many women, of eligible age, do it every year”, says Laura Testa.
The problem goes beyond the availability of equipment, but the absence of an organized structure. “We still have what we call opportunistic screening — people do it when they can. In fact, we don’t have this organized in the private system either”, says the oncologist.
Fernando Maluf sees AI as a tool to reduce inequalities. For him, it would be possible to generally replace the radiological reading of the exam, as long as applications and technological systems trained for the function and validated are used, which would be welcome in “areas with fewer radiologists available or where there is more deficient training of radiologists”.
This horizon of change extends beyond mammography. “We have studies today, for example, of MRI in prostate cancer with similar results, as well as studies in lung nodules for diagnosing early lung cancer”, says the doctor.