Development of AI algorithms in BreastScreen Norway

Background

Breast cancer is the most common cancer among women in Norway and globally. Although some risk factors for breast cancer are known, the disease is difficult to prevent for the individual woman. Early detection through screening is therefore considered an effective tool in reducing mortality from the disease.

A standard screening examination in BreastScreen Norway consists of mammograms taken from two angles of each breast. All examinations are assessed by two radiologists, both interpreting the images without knowing what the other has concluded (independent double reading). However, very few women participating in the screening program have signs of breast cancer, and radiologists therefore spend a lot of time assessing normal mammograms.

With recent advances in artificial intelligence (AI), there is a potential to improve the current screening programme, for example by using AI as a support for the radiologists in their assessment of mammograms. By AI we here refer to digital systems that learn to recognize patterns in mammograms that may be signs of breast cancer, because they have been trained to do so by analysing large amounts of data over time.

BreastScreen Norway has two ongoing research projects that use AI to develop new algorithms for assessing mammograms; "Machine learning In Mammographic screening" (the MIM study) and "AIforScreening; Robust and reliable AI for breast cancer screening" (AIforScreening). The Norwegian Computing Center is responsible for developing the algorithms.

Purpose

The overall aim of the two projects is to develop AI algorithms that can be used as a support for image assessment in BreastScreen Norway.

The main aim of the MIM study is to develop a digital system that can select images with a high probability of not having signs of breast cancer. This can streamline and increase the quality of BreastScreen Norway, among other things by allowing radiologists to spend more time on women who have signs of breast cancer.

AIforScreening is a continuation of the MIM study aiming to further develop the AI algorithms from the first study to become even more robust and reliable. The goal is that the algorithms can be used on mammograms acquired from different vendors, and that they will be able to use the information from prior screening examinations in the assessment of current images.

Data

In order to develop good and accurate algorithms adapted to BreastScreen Norway, large amounts of image data from screening examinations are required. In addition, information related to attendance and findings on the mammograms is used, such as the radiologists’ assessments and the results of the screening examinations (screening information).

The MIM study is based on data from screening examinations conducted in BreastScreen Norway at seven health trusts within the Northern Norway Regional Health Authority, the Central Norway Health Authority and the South-Eastern Norway Regional Health Authority. The examinations have been conducted in the period from the introduction of digital mammography at each health trust (between 2004 and 2011) onwards.

In AIforScreening, the MIM data will be reused. In addition, image data and screening information will be collected from three additional health trusts in the South-Eastern Norway Regional Health Authority.

The project group at the Cancer Registry of Norway will collect image data from the collaborating health trusts and screening information from the Cancer Registry's databases. The data will be pseudonymised and quality assured by the project group before being delivered to the Norwegian Computing Center, which will develop and train the AI algorithms.

The projects only use data from women who have not opted out of having their information related to screening examinations with negative findings stored in the Cancer Registry, in accordance with the Cancer Registry Regulations. The women will not be contacted about the projects, and it will be impossible to recognize individuals in published results.

Organization

The Cancer Registry of Norway is project manager and data controller, and has the overall administrative responsibility for the projects. Among other things, we will ensure that the projects have all the necessary approvals, collect image data and screening information in the project, test the AI algorithms and draft any implementation plans in BreastScreen Norway.

The Norwegian Computing Center is data processor in the projects. They have high professional expertise in image analysis and AI, and are responsible for developing the AI algorithms.

The health trusts, represented by the breast centres, are partners in the projects. They are the regional specialists in breast cancer screening and diagnostics, and will assist with radiologic expertise and knowledge of practical screening, as well as facilitate the extraction of image data.

The University of Tromsø has special expertise in IT systems for biological and medical applications, and will act as an important advisor on aspects related to this. They will also supervise master's students on related subjects.

Karolinska Institutet will contribute with their expertise on AI in mammography.

Status as of May 2023 

In the MIM study, new algorithms have been developed as planned. The algorithms have also been tested on  "new" datasets, i.e. a dataset that was not part of the data in the development phase. Preliminary analyses suggest that the AI algorithms have the potential to increase the sensitivity of mammography screening by increasing the number of screen-detetcted breast cancers, reducing the number of interval cancers and reducing the workload of the radiologists.

The project has been completed in the Research Council of Norway and no longer has external funding. The project completion date in Regional Commitee for Medical and Health Research Ethics is extended due to the need of further developing, validating and improving the algorithms. This work is ongoing.

In AIforScreening, necessary approvals are obtained and agreements signed. Since a large amount of new image data (about 650,000 examinations with 2.6 million mammograms) will be added to the existing dataset from the MIM study, the data collection is time-consuming. Extensive work is underway at the Cancer Registry to facilitate and implement this, in collaboration with external IT resources. Until the new dataset is available, the Norwegian Computing Center uses the information from the MIM study to work with the project's objectives.