Automated measurements of breast segmentation are becoming an essential process to the development of quantitative and reproducible imaging biomarkers. Background parenchymal enhancement has been suggested as a possible biomarker to assess breast cancer risk and for early prediction of neoadjuvant response. Manual processing is time consuming, prone to inter- and intra-rater reliability limitations, and therefore not practical for use as a regular clinical metric or for large-scale data imaging research. Background parenchymal enhancement represents the mean value of the initial enhancement of normal fibroglandular tissue, and because it represents an averaged value, requires only a large portion of tissue be segmented for accurate measurement. We have tested a method for automated breast tissue segmentation and assessed its performance using a pilot dataset.

In the IRB/HIPAA compliant study, breast tissue segmentation (Figure 1) was obtained using a coronal reformat of a pre-contrast MRI axial series resampled to a 1.0 mm³ isotropic voxel size. Image processing algorithms were applied using MATLAB and Statistics Toolbox Release 2012a (The MathWorks, Inc., Natick, Massachusetts, United States). Parallel imaging artifacts were eliminated and an initial tissue/air mask was obtained. The mask was traversed slice by slice from anterior to posterior and the number of 2D connected components in each slice was tracked. When the number of connected components dropped from 2 to 1 signaling that the sternal region was reached, an additional 1 cm band in the A/P direction was included and mask slices beyond that point were eliminated. The mask was then resampled into the original image matrix and b-splines were applied for manual editing using in-house software programmed in IDL (ITT Visual Information Solutions, Boulder, CO, USA).

The automated breast segmentation method was then applied to a test dataset of 49 patients with bilateral axial breast MRIs. These examinations were originally obtained in patients who received breast MRI to monitor neoadjuvant response. Two reviewers evaluated the segmented examinations qualitatively against the reference standard of manually drawn breast contours. Method accuracy was assessed if manual modification was required based on volume of breast tissue covered and extent of skin/chest wall inclusion to be sufficient for background parenchymal enhancement measurement.

Of the 49 examinations tested, 61% of patients had fibroglandular tissue successfully segmented that required no further modification in order to measure background parenchymal enhancement (e.g. Figure 2a). A further 13% had satisfactory breast segmentation that only required minimal manual modification. The final 26% required significant manual adjustment either due to segmenting <50% of breast tissue or including skin/chest wall (e.g. Figure 2b). Of those that required significant manual adjustment, most were due to inadequate breast volume covered and required adjustment of the b-splines to extend posteriorly to the chest wall. A single case with breast reconstruction and implants required the most significant adjustment. Overall, segmentation cases captured more of the extreme craniocaudal sections of the breast than was obtained by manual segmentation, but performed worse along the posterior segmentation boundary.

Initial results suggests that this automated segmentation method was successful in approximately 75% of bilateral breast MRI examinations in our pilot dataset requiring minimal to no manual correction. The method in its current form would likely characterize background parenchymal enhancement accurately, and improved upon manual segmentation by including the extreme craniocaudal sections of the breast. Future versions will focus on better characterization of the breast-chest wall boundary and optimization with non-symmetric breasts such as post-reconstruction patients.