Automated rating of background parenchymal enhancement in MRI of extremely dense breasts without compromising the association with breast cancer in the DENSE trial

Hui Wang; Bas H M van der Velden; Erik Verburg; Marije F Bakker; Ruud M Pijnappel; Wouter B Veldhuis; Carla H van Gils; Kenneth G A Gilhuijs

doi:10.1016/j.ejrad.2024.111442

Automated rating of background parenchymal enhancement in MRI of extremely dense breasts without compromising the association with breast cancer in the DENSE trial

Eur J Radiol. 2024 Mar 24:175:111442. doi: 10.1016/j.ejrad.2024.111442. Online ahead of print.

Authors

Hui Wang¹, Bas H M van der Velden¹, Erik Verburg¹, Marije F Bakker², Ruud M Pijnappel³, Wouter B Veldhuis³, Carla H van Gils², Kenneth G A Gilhuijs⁴

Affiliations

¹ Image Sciences Institute, University Medical Center Utrecht, Utrecht, The Netherlands.
² Julius Center for Health Sciences and Primary Care, Utrecht, The Netherlands.
³ Department of Radiology, University Medical Center Utrecht, The Netherlands.
⁴ Image Sciences Institute, University Medical Center Utrecht, Utrecht, The Netherlands. Electronic address: K.G.A.Gilhuijs@umcutrecht.nl.

PMID: 38583349
DOI: 10.1016/j.ejrad.2024.111442

Abstract

Objectives: Background parenchymal enhancement (BPE) on dynamic contrast-enhanced MRI (DCE-MRI) as rated by radiologists is subject to inter- and intrareader variability. We aim to automate BPE category from DCE-MRI.

Methods: This study represents a secondary analysis of the Dense Tissue and Early Breast Neoplasm Screening trial. 4553 women with extremely dense breasts who received supplemental breast MRI screening in eight hospitals were included. Minimal, mild, moderate and marked BPE rated by radiologists were used as reference. Fifteen quantitative MRI features of the fibroglandular tissue were extracted to predict BPE using Random Forest, Naïve Bayes, and KNN classifiers. Majority voting was used to combine the predictions. Internal-external validation was used for training and validation. The inverse-variance weighted mean accuracy was used to express mean performance across the eight hospitals. Cox regression was used to verify non inferiority of the association between automated rating and breast cancer occurrence compared to the association for manual rating.

Results: The accuracy of majority voting ranged between 0.56 and 0.84 across the eight hospitals. The weighted mean prediction accuracy for the four BPE categories was 0.76. The hazard ratio (HR) of BPE for breast cancer occurrence was comparable between automated rating and manual rating (HR = 2.12 versus HR = 1.97, P = 0.65 for mild/moderate/marked BPE relative to minimal BPE).

Conclusion: It is feasible to rate BPE automatically in DCE-MRI of women with extremely dense breasts without compromising the underlying association between BPE and breast cancer occurrence. The accuracy for minimal BPE is superior to that for other BPE categories.

Keywords: Background Parenchymal Enhancement; Breast MRI; Dense breasts; Machine learning.