Mirai: estimating breast cancer risk from a mammogram

Why estimate breast cancer risk?

Today, breast cancer screening often follows the same rules for all women in an age group. The authors point out that better risk models would allow screening to be targeted: finding earlier the people who need it and reducing the downsides of screening for others.

Standard tools, such as the Tyrer-Cuzick model, combine health information and family history. The idea in this study is to make use of the mammogram itself, which holds information the human eye may not pick up.

What did the researchers do?

They built Mirai, a deep learning model: a program that learns to recognise patterns by looking at a very large number of images. Mirai was trained on a large dataset from Massachusetts General Hospital (MGH) in the United States.

It was designed to do three things: estimate risk at several time points, make use of risk information that is sometimes missing (such as family history), and give consistent estimates from one mammography machine to another. It was then tested on held-out datasets from MGH, Karolinska University Hospital in Sweden and Chang Gung Memorial Hospital in Taiwan.

What were the results?

The authors measure performance with the C-index, a score showing whether the model ranks women correctly according to whether they develop cancer (0.5 is chance, 1 is a perfect ranking). Mirai scored 0.76 at MGH, 0.81 at Karolinska and 0.79 at Chang Gung.

The authors also report that Mirai obtained significantly higher 5-year ROC AUCs than the Tyrer-Cuzick model. Results were therefore good not only in the original hospital, but also in different populations and with different equipment.

What could this mean for women?

If such models are validated in practice, they could help offer closer or more spaced screening, or extra tests, depending on each woman's estimated risk. This might allow some cancers to be found earlier and spare other women unnecessary tests.

For now, Mirai remains research. A risk score is not a diagnosis: it does not say a woman will get cancer, only that her risk is higher or lower.

Because mammograms are already taken in screening, an estimate from the same image would not require a new test.

What are the limits of this study?

The study is retrospective: it uses mammograms already taken, for which the outcome is known. It does not yet show that using Mirai in clinic improves women's health or real-life screening.

The three hospitals tested do not represent every population or every piece of equipment. The authors stress that the model's accuracy still needs refining and that it must be shown to improve workflows before clinical use.