Authors: Taylor Forbes, Rachel Harris, Benjamin Jackson, Nicole Peterson, Sydney Tanner, Chloe Nelson
Mentors: Christopher Dillon
Insitution: Brigham Young University
Focused ultrasound (FUS) therapy is a non-invasive therapy for breast cancer. Treatment plans for this therapy are created on a patient-to-patient basis, which requires a significant amount of time from medical professionals. An important and time-consuming part of developing treatment plans is the precise segmentation of the breast magnetic resonance imaging (MRI) scan and subsequent treatment simulation to ensure that the treatment is effective and safe. Segmentation involves dividing the MRI dataset into segments by assigning distinct tissue types that are then assigned properties and used in simulations to help clinicians plan FUS treatments. However, imprecise interfaces between different tissue types in MRI images lead to discrepancies between individual segmentations, thereby introducing variability into the segmentation process. This variability—which is found even among expertly performed segmentations—can lead to differences in treatment plans.
Here, analysis was performed in order to quantify interobserver variability in breast MRI segmentation. This study was conducted by providing basic segmentation training to undergraduate research assistants with no prior segmentation experience. Each participant segmented the same breast MRI dataset into different tissue types using the software Seg3D. The different segmentations were then compared using contour similarity metrics (such as the Dice Similarity Coefficient and Jaccard Index) as well as tissue volume differences. The interobserver variability was quantified using the results from these analyses, which will be helpful in determining the level of precision required for the use of a given segmentation in FUS treatment planning.