Building the Malmö Breast Imaging Database for Evaluation of Artificial Intelligence
Technology development pushes us to find new ways of working, also in breast screening. Breast screening has been organized in the same manner since the 1980's, apart from the switch from analogue to digital mammography (DM) in late 1990's, which was a huge practical leap forward. In Sweden and many northern European countries double-reading is practiced, where two breast radiologists independently read the images, in order to increase both sensitivity and specificity. The first CAD (computer-aided detection) systems had issues with specificity (1). With the introduction of deep-learning and neural networks, the algorithms used today may be "self-improved" and there is a great chance that artificial intelligence (AI) will change breast imaging used either as a stand-alone tool for e.g. risk prediction or in conjunction with the human reader as a decision aid. Furthermore, if digital breast tomosynthesis (DBT) is to be used in screening, reading times will be prolonged (approximately doubled). In addition, there is a lack of breast radiologists.
Background: Only in Sweden, approx. 700K women 40-74 years of age, attend mammography screening every year; making up one of the largest, long-term preventive projects in female population health. Digital breast tomosynthesis, DBT is a form of 3D-mammography that overcomes the overlapping tissue effect inherent in mammography and hence detects more breast cancers. Internationally, the Malmö Breast Tomosynthesis Screening Trial, MBTST, is one of the largest prospective trials and provides essential data forming the evidence-base for tomosynthesis in screening. The results from our trial recently published in The Lancet Oncology show a 34% increase in cancer detection (2). In parallel, we have through collaboration with Dutch researchers investigated the use of artificial intelligence (AI) and machine learning (ML) in breast screening by using an international retrospective cancer enriched cohort. The results were recently published in JNCI (3) and another manuscript accepted in Eur Radiol (A Rodriquez-Ruiz, et al). Since it is important to assess AI in more realistic screening settings, we subsequently performed analyses using a subcohort of the DM cases in our screening trial MBTST. We were able to prove that stand-alone AI could improve the efficiency in breast cancer screening by safely excluding one-fifth of normal mammograms, including a small fraction of false positives, from screen reading performed by radiologists without missing cancers. This may consequently reduce work load and costs in screening (K Lång et al. submitted).
Aim: The overall aim of this research project is to build large imaging databases to be used for development and evaluation of AI and machine learning software. The following specific aims will be investigated:
1) Development and validation of machine learning in breast cancer screening
a. To study the accuracy of ML compared to human readers in digital mammography
b. To study the accuracy of ML compared to human readers in digital breast tomosynthesis
c. Assess if ML can improve the screening workflow by discarding clearly normal screening examinations
d. Assess the feasibility of using ML to reduce the number of interval cancers
2) Development of radiological breast cancer risk profiles
a. Development of radiological risk profiles based on breast density, texture analysis, and image findings
b. Evaluation of potential changes in breast density and texture analysis over a longer period of time vs. tumour evolvement and growth
The already completed image and cancer database from the MBTST (please see CAN 2016/488, projects 1a, 1b, 2a and my application for continued project grant support 2020-22) guarantees the feasibility of the proposed projects, to be built on to expand the database for future analyses.