BADDI: Artificial intelligence in screening with standard mammography and tomosynthesis
Background
BADDI stands for Breast Cancer, Artificial Intelligence, Digital Breast Tomosynthesis, Digital Mammography and Interval Cancer.
Artificial intelligence and machine learning have shown promising results in several areas of mammographic screening. However, much is still unknown about the advantages and disadvantages of such use, and there are few studies on machine learning in screening, and especially with the use of tomosynthesis. Tomosynthesis is a 3D-like and more advanced type of mammography than standard digital mammography. The To-Be studies - the Tomosynthesis studies in Bergen - have investigated whether tomosynthesis is a suitable screening technique in BreastScreen Norway. The BADDI-project is based on data from these studies.
Interval cancers are breast cancers detected between two screening tests. They are an inevitable part of a screening program. In radiological review studies from, among others, BreastScreen Norway, radiologists have re-assessed diagnostic images together with the screening mammograms that preceded the women's interval cancer. These studies have found that 1 in 4 interval cancers were identifiable on previous screening mammograms, when the radiologists were informed about the diagnosis and knew where to look for the tumour. These breast cancers were classified as "missed", while the rest were classified as "true". In screening programs, it is desirable to keep the proportion of interval cancer cases low, and it will be interesting to investigate whether machine learning systems can detect breast cancers that are particularly difficult to detect for radiologists, both the "missed" and "true" interval cancer cases.
Purpose
The purpose of the BADDI-project is to increase our knowledge about the ability machine learning has to detect breast cancers in screening with tomosynthesis and standard digital mammography, and to investigate whether machine learning systems can be as good or better at detecting breast cancers than radiologists.
This project will increase our knowledge about what type of tumours machine learning systems can detect - are they the same tumours as radiologists find, or can the machines also detect breast cancers that the human eye does not perceive? The project will also provide knowledge about whether machine learning systems can detect more breast cancer cases without simultaneously increasing the proportion of false positive screening tests, and without finding more small, slow-growing breast cancer tumours that can increase the risk of overdiagnosis and overtreatment of women.
Data
BADDI will use information from screening examinations already conducted in the To-Be studies that were carried out in BreastScreen Norway in the period 2016-2019. We will use information from the mammography and tomosynthesis images (image data), the radiologists' interpretations, and results from the screening examinations (screening information), in line with the Cancer Registry Regulations.
In the To-Be studies, about 500 breast cancer cases were detected (screen-detected and interval cancers). These cases, in addition to randomly selected screening examinations from women who had a negative or false positive screening result, form the basis of the study population in BADDI.
BADDI is a register-based study, and the women will not be contacted again about the project. It will be impossible to recognize individuals in published results.
Planned studies and status
Study 1: Machine learning versus radiologists - use of machine learning to interpret mammograms from women with breast cancer.
The goal is to allow a machine learning system to interpret screening mammograms from To-Be 1 and To-Be 2 to examine the sensitivity and specificity of the system. We want to explore whether 1) the machine learning system detects the same screen-detected breast cancers as the radiologists, and 2) whether the machine-learning system is able to detect the interval cancers from To-Be 1, and the screen-detected breast cancers from To-Be 2 on the screening mammograms from To-Be 1. The results will be compared for tomosynthesis and digital mammography.
Status: As of December 2022, we have retrieved image data from women who were screened with digital mammography, and the machine learning system has analyzed these mammograms. The system is currently unable to analyze tomosynthesis images, but we expect to solve this early 2023.
Study 2: Use of machine learning to investigate whether the number of "missed" breast cancers can be reduced.
Prior to BADDI, a radiological review were performed of mammograms from women screened with tomosynthesis in To-Be 1 and diagnosed with interval breast cancers after screening in To-Be 1 or with screen-detected cancer in To-Be 2. The cases were classified as "missed" or "true" in To-Be 1.
The aim of study 2 is to investigate the machine learning system's sensitivity for detecting "missed" and "true" interval and screen-detected breast cancers in tomosynthesis images. We want to explore whether the machine learning system, by interpreting the tomosythesis images from screening prior to diagnosis of interval cancer in To-Be 1 and screen-detected breast cancer in To-Be 2, is able to detect breast cancers that have been classified as "missed" or "true" in our radiological review.
Status: As of December 2022, the radiological review has been carried out (May 2022). The review was carried out by external radiologists, which means radiologists working at other Norwegian breast centers and internationally. We are now waiting for an updated version of the machine leraning system, to be able to analyze tomosynthesis images. This is expected early 2023. The results from the review will be compared with the results from the machine learning system's interpretation of the tomosynthesis images.
Organization
BADDI is a collaborative project between the Cancer Registry of Norway, the breast center at Haukeland University Hospital and the Mohn Medical Imaging and Visualization Center at Haukeland University Hospital (MMIV). The Cancer Registry is data controller. We are responsible for obtaining the necessary approvals, testing the machine learning system and extracting information about the women in the study population, and transferring this to MMIV. The Cancer Registry will also be responsible for performing analyzes, and interpreting and publishing the results.
MMIV at Haukeland University Hospital is responsible for extracting, pseudonymizing and analyzing image data using the machine learning system. They will also contribute in the interpretation and publication of results.
ScreenPoint Medical B.V. has developed the machine learning system to be used in this study, Transpara ™. The system is FDA approved. ScreenPoint Medical is responsible for access, installation, calibration and training the staff in how to use the system.