Why is screening mammography not perfect?
Screening mammography is a breast X-ray offered to women without symptoms, to find a cancer early, when treatment is more likely to succeed. Screening programmes exist in many countries.
But reading these images is hard. There are false positives, meaning false alarms that lead to extra tests and anxiety, and false negatives, meaning cancers that go unnoticed. In the United Kingdom, each mammogram is read by two radiologists, which takes a lot of staff.
A reliable AI could therefore help both women (fewer errors) and services (less reading workload).
What did the researchers do?
An international team developed an AI and evaluated it on two large data sets: a UK set, representative of the women screened, and a larger US set enriched with cancer cases. The authors also checked whether an AI trained on UK data worked on US data.
They then compared the AI with six radiologists in an independent study, and simulated its role in the double reading used in the UK.
Testing the AI in two countries matters: a tool that only works in the hospital where it was designed would be of little use.
What results were obtained?
Compared with the doctors' readings, the AI cut false positives by 5.7 points in the US and 1.2 points in the UK, and false negatives by 9.4 points in the US and 2.7 points in the UK, in absolute terms. The results carried over from one country to the other.
In the study with six radiologists, the AI did better than every one of them: its area under the curve was higher than that of the average radiologist by 11.5 points. In the simulation where it takes the place of one of the two readers, it kept non-inferior performance compared with the current system and cut the second reader's workload by 88%.
In other words, in this evaluation the AI made fewer mistakes of both kinds than the doctors whose readings served as the comparison.
Key figures
- 5.7 and 1.2 pointsfewer false positives in the US and the UK (9.4 and 2.7 points fewer false negatives)
- 11.5 pointshigher area under the curve than the average radiologist, in the six-radiologist study
- 88%less work for the second reader, in a double-reading simulation
What could this mean for patients?
Fewer false alarms means fewer needless tests and less worry. Fewer missed cancers means earlier diagnoses. And lighter reading loads could free radiologists' time, which is precious in countries that lack them.
The authors see these results as a basis for launching clinical trials aimed at improving the accuracy and efficiency of breast cancer screening.
Because many women are screened, even small gains in accuracy could affect a large number of people.
What are the limits of this study?
It is an evaluation on images that had already been read, not a trial in which the AI was used live. We therefore do not know whether it truly improves care for women. The US set was enriched with cancer cases, which is not the situation in real screening.
The six-radiologist study is small, and the double-reading simulation is not the same as operating in real conditions. The clinical trials the authors call for are therefore needed.
Finally, as in any evaluation on past records, longer follow-up of the women could change some of the results.
Preparing for your next mammogram with Mima
In this study, an AI reading mammograms reduced false alarms and missed cancers. It is not everywhere yet: some centres are trying these tools, others are not, and a radiologist is still involved.
Mima lets you note your mammogram appointments and their results. You can also prepare the question: "Does my centre use AI to review the images?" And if an alert worries you, she helps you gather your questions for the consultation.
Frequently asked questions
What is a false positive in mammography?
It is a false alarm: the image looks suspicious when there is no cancer. It often leads to extra tests and anxiety.
Does this AI replace the radiologist?
Not in this study. The authors simulated its role as one of the two readers in a double-reading system. A human stays involved.
Why does the study use data from two countries?
To check that the AI works beyond the country where it was trained. The authors show that it carries over from the UK to the US.
Is the AI already used in screening programmes?
The study is an evaluation on existing images. It prepares the ground for clinical trials but does not yet prove a benefit for patients.
The original study
McKinney SM, Sieniek M, Godbole V, et al. “International evaluation of an AI system for breast cancer screening” Nature, January 2020. DOI : 10.1038/s41586-019-1799-6
Read the full study in NatureThis article summarises a scientific study and is not medical advice.