Data¶

The PRISM-AI Challenge evaluates mammography-based AI models using data from the Personalized RIsk-based MAmmography Screening study (PRISMA).

The PRISMA dataset is hidden: no images or outcome data are distributed to participants. Algorithms are executed on the Grand Challenge platform, where they process the images and produce predictions that are evaluated automatically.

The PRISMA study¶

PRISMA is an observational, population-based, prospective cohort study, nested within the Dutch breast cancer screening programme. Women aged 50–75 invited for breast cancer screening at 20 screening units across the Dutch regions East, North, South and South-West were recruited between 2014 and 2019, yielding 78,736 participants.

Follow-up information on incident breast cancer is available through linkage with the Netherlands Cancer Registry (NCR), which also provides clinicopathologic data and the month and year of diagnosis.

Which examination is used¶

For each participant, the Challenge uses the mammogram taken at study entry. Each examination consists of the four standard "for presentation" views:

Inclusion and exclusion criteria¶

The Challenge evaluates algorithms as they would be used for risk assessment in personalized screening. That assessment takes place on the basis of a negative screening examination, in order to decide on a subsequent personalized screening strategy. Women who were recalled at the examination used for prediction are therefore excluded. Furthermore, it is required that all four "for presentation" are available and that no breast implant is present. Figure 1 shows the flowchart for the primary analyses of the PRISM-AI Challenge, in which the mammograms of 72,076 PRISMA participants will be used.

Outcome data¶

The primary outcome is 5-year breast cancer risk, including both invasive breast cancer and ductal carcinoma in situ (DCIS).

Within the 72,076 included participants, the linkage that is currently available identified 966 breast cancers after a median follow-up of 4.3 years (IQR 4.0–4.6).

The registry linkage has been updated at the start of the Challenge, identifying additional cases and extending follow-up by approximately 2.5 years. This allows accurate estimation of 5-year breast cancer risk.

Data format¶

Mammograms are provided to your algorithm on Grand Challenge as stacked images in MetaImage (.mha) format, with the relevant DICOM header information supplied as accompanying metadata in .json format.

A demo inference container is available on Github. A small set of example mammograms are provided to participants in the exact format used for evaluation.