To compare magnetic resonance imaging (MRI) derived breast density measurements using automatic segmentation algorithms with radiologist estimations using the Breast Imaging Reporting and Data Systems (BI-RADS) density classification.
Materials and Methods
40 women undergoing mammography and dynamic breast MRI as part of their clinical management were recruited. Fat-water separated MR images derived from a 2-point Dixon technique, phase sensitive reconstruction and atlas based segmentation were obtained before and after intravenous contrast. Breast density was assessed using software from Advanced MR Analytics (AMRA), Linköping, Sweden with results compared to the widely used four-quartile quantitative BIRADS scale.
The proportion of glandular tissue of the breast on MRI was derived from the AMRA sequence. The mean unenhanced breast density was 0.31 ± 0.22 (mean ± SD) (left)
and 0.29 ± 0.21 (right). Mean breast density on post-contrast images was 0.32 ± 0.19 (left) and 0.32 ± 0.2 (right). There was "almost perfect" correlation between pre and post-contrast breast density quantification: Spearman correlation rho=0.98 (95% confidence intervals (CI): 0.97-0.99) (left) and rho=0.99(CI: 0.98-0.99) (right). The 95% limits of agreement were -0.11-0.08 (left) and -0.08-0.03 (right).
Interobserver reliability for BIRADS is "substantial": weighted Kappa k=0.8 (CI: 0.74- 0.87). The Spearman Correlation coefficient between BIRADs and MR breast density was rho=0.73 (CI: 0.60-0.82) (left) and rho=0.75 (CI: 0.63-0.83) (right) which is also "substantial".
The AMRA sequence provides a fully automated, reproducible, objective assessment of fibroglandular breast tissue proportion that correlates well with mammographic assessment of breast density with the added advantage of avoidance of ionising radiation.