Mammography

Detect breast cancer from mammograms

10 participants · own branch each

Why this challenge exists

A radiologist reads a mammogram in under a minute, several hundred times a week.

Screening works because it is done at scale, and it is fragile for the same reason: the lesions that get missed are rarely the obvious ones, they are the ones read at the end of a long list. A second reader halves that risk, and there are not enough second readers.

A model that flags the images worth a second look does not replace anyone. It changes the order in which the list is read — which is the part of the process nobody currently controls.

Who hosts this challenge

MyTwin Lab

The brief

Objective

What has to be shipped:

A detection model trained on the challenge's annotated mammograms, scoring each image rather than sorting it.
A threshold analysis: what the model costs in false positives at the sensitivity screening actually requires.
A view of the flagged region, so a radiologist can see what the model saw instead of taking its word.

Expected result

A pull request from your branch, with the board reflecting what it contains.
Measured sensitivity and specificity against the reference set, reported at the operating point you recommend and at two others.