Trustworthy AI for breast cancer screenings

We are developing machine learning models to further improve and streamline the national breast screening programme.

Breast cancer is the most common type of cancer among women in Norway. Early detection is key in order to limit the severe cases of the disease and ensure higher survival rates. Because of this, all women in Norway aged 50 to 69 are invited to participate in a biennial x-ray examination through the national screening programme, BreastScreen Norway.

Here, images are taken of both breasts and subsequently examined by two independent radiologists who specialise in mammography. As less than 1% of mammograms show signs of cancer, doctors spend a lot of time examining images of healthy people.

Technological advancements have made it possible to automise parts of the screening process, and current research shows that machine learning models are just as accurate as human experts. Automisation frees up time and resources, making it possible for doctors to focus on patients in need of further examination and those who are actually ill.

This is an x-ray image of a breast. Breast tissue is visible in the breast and a possible cancerous area is located with a red dot and a red square around it. — *Caption*: A two-stage detection method. The first model identifies the area inside the red square as suspicious, and the second model locates the cancerous area, marked with a red dot. Image: NR/The Cancer Registry of Norway

Adaptable models for different equipment and recording conditions

We are developing methods and models for AI-based analysis of mammograms in collaboration with the Cancer Registry of Norway.

In our previous project, Machine Learning in BreastScreen Norway (MIM), we developed a prototype of a model for AI-based analysis of mammograms using a large image dataset from the Cancer Registry of Norway. During training, the model only received information about which images showed signs of cancer, but had to localise relevant areas of the breast without any further direction.

The objective of AIforScreening is not only to continue our previous research in detection, but to also examine how the models can be made more robust, as equipment and recording conditions change over time. In addition, we are studying how the contribution of time series can ensure safer identification, and how AI can be integrated in a way that facilitates productive interactions between radiologists and the models.

A two-stage method for accurate detection

It is crucial that the models aren’t so-called black boxes generating conclusions without transparency. The models should also allow for a degree of uncertainty and be able to localise possible signs of breast cancer. This is beneficial in two notable ways. Firstly, the models become more transparent and understandable. Secondly, the models themselves benefit from the explanations.

We have developed two cooperating models, where the first identifies the area of the image that most resembles cancer, and the other model zooms in and examines the area more closely. Our two-stage method has been presented at the NORA conference and our results are published in the scientific journal Nordic Machine Intelligence. Our model showed results that were on par with a radiologist, and outperformed the commercial product we compared it with.

Fredrik Andreas Dahl

Senior Reseach Scientist

Project: AIforScreening: Robust and trustworthy AI for breast cancer screening

Partners: The Cancer Registry of Norway

Funding: The Research Council of Norway

Period: 2021 – 2025

Further reading:

Dahl, F. A., Brautaset, O., Holden, M., Eikvil, L., Larsen, M., & Hofvind, S. (2023). A two-stage mammography classification model using explainable-AI for ROI detection. Nordic Machine Intelligence, 3, 1-7. https://doi.org/10.5617/nmi.10459

Machine learning in BreastScreen Norway (MIM) – project page

BreastScreen Norway – kreftregisteret.no