Assessment of breast composition in MRI using artificial intelligence e A systematic review

Introduction: Magnetic Resonance Imaging (MRI) performs a critical role in breast cancer diagnosis, especially for high-risk populations e.g. family history. MRI could take advantage of the implementation of artificial intelligence (AI). AI assessment of breast composition factors is less studied than those of lesion detection and classification. These factors are breast density, background parenchymal enhancement (BPE) and fibroglandular tissue (FGT), which are recognized breast cancer phenotypes. Methods: Following PRISMA guidelines, the PROSPERO registered review examined the role of AI in assessing breast composition in MRI. A search of articles from Pubmed, Ovid, Embase, Web of Science, Cochrane, and Google scholar from 2010 to 2022 was conducted. Peer-reviewed, in-vivo studies were included based on defined search categories. Adapted QUADAS-2, CASP and Covidence tools were utilized for quality assessment. Results: Seven studies were identified as being of sufficiently high quality. The studies showed that AI has the potential to provide a comparable level of accuracy against the relevant reference standard. There were limited performance results when delineating BPE and FGT BI-RADs categories. The review highlighted the variability in AI models while the range of statistical methods and small cohort sizes limited cross study compatibility. Conclusions: AI has potential in assessing breast composition in MRI. However, variability in AI systems deployed and statistical measurements alongside limited validation across diverse populations remain an issue. AI systems may perform better with binary categorizations rather than the quaternary spectrum of BI-RADS. Implications for practice: AI could assist in developing personalized breast composition assessments. Future developments could focus on better delineation of breast composition categories. AI models that have trained on more diverse and larger populations should result in more robust and effective clinical applications. (c) 2025 The Authors. Published by Elsevier Ltd on behalf of The College of Radiographers. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).