Challenge Banner

Overview & Goals

The PRISM-AI Challenge is an independent, transparent benchmark of AI algorithms for breast cancer risk prediction, evaluated on a large population-based screening cohort of over 72,000 women.

Status: The Challenge has moved into the phase in which the submitted algorithms will be run on the hidden PRISMA set and evaluated. Results will be published on these pages.

AI for personalized breast cancer screening

In the Netherlands, all women aged 50 to 75 at "average" risk are invited for mammography screening every two to three years. However, breast cancer risk varies widely within this population denoted as "average", and a one-size-fits-all approach does not address these differences.

AI-driven personalized screening could be the next major step forward: accurate risk estimates could help tailor screening intensity, improving early detection for some while reducing unnecessary procedures for others. Several commercial and academic AI algorithms now predict future breast cancer risk from screening mammograms.

The PRISM-AI Challenge

The PRISM-AI Challenge provides a transparent, independent, and rigorous evaluation of AI algorithms for breast cancer risk prediction on a large population-based screening cohort. This Challenge uses data from the Personalized RIsk-based MAmmography screening study (PRISMA). The PRISMA study is an observational, population-based, prospective cohort study, nested within the Dutch breast cancer screening program.

PRISM-AI is a benchmark of existing algorithms, not a model development competition. There is no development phase and no training data are distributed. Participating teams submit an algorithm they have already trained, as an inference container on the Grand Challenge platform. Algorithms are executed on Grand Challenge against the hidden PRISMA set, and performance metrics are computed automatically using pre-defined and publicly available evaluation scripts.

What is evaluated

The primary outcome is the probability of being diagnosed with breast cancer within 5 years of the screening examination, including both invasive breast cancer and ductal carcinoma in situ (DCIS). For algorithms that output a 5-year risk estimate, we evaluate:

  • Risk distributions: the distribution of predicted risks for cases and non-cases
  • Discrimination: how well the algorithm separates women who do and do not develop breast cancer
  • Calibration: whether predicted risks match observed risks
  • Clinical utility: net benefit across relevant decision thresholds

These results are published on the leaderboards and pages of this platform. See Evaluation for definitions of every metric. All evaluation scripts are available on GitHub.

Eligible algorithms

The Challenge evaluates algorithms that predict future breast cancer risk from screening mammograms and output a breast cancer probability score for a specified time period.

The PRISM-AI project

The Challenge is one part of the wider PRISM-AI research project. Alongside the benchmark, the project uses the resulting risk predictions to determine which risk thresholds and screening strategies offer the best balance between the benefits and harms of screening, and investigates which forms of personalized screening women themselves consider acceptable. PRISM-AI is funded by ZonMw and runs from 2025 to 2027; a project summary is available on the ZonMw project page.