Author(s): Jacob Lehnhof, Taylor Forbes, Nicole Peterson, Rachel Harris, Ben Jackson, Jacob Moulder, Drew Wagstaff, Amelia Benedict, Brianne Geiger, Estee Revill
Mentor(s): Christopher Dillon
Institution BYU
Magnetic resonance-guided focused ultrasound (MRgFUS) is a non-invasive treatment that uses ultrasound waves to deposit thermal energy and ablate tumors in situations such as breast cancer. Currently, clinicians plan treatments by relying on their experience and intuition and use magnetic resonance imaging (MRI) guidance to make real-time adjustments during MRgFUS treatment. In the future, pre-treatment simulations could be used to enhance pre-treatment planning and reduce the need for real-time adjustments, improving the patient’s experience and treatment outcomes. One important input for MRgFUS simulations is a 3D model of the region of interest - in this case, the patient’s breast. Observers create this 3D model, or segmentation, by labeling each voxel as a specific tissue in a MRI dataset. Resolution limitations and imprecise interfaces between tissue types cause ambiguity regarding the tissue type of each voxel and lead to variations between segmentations as individual observers, including experts, interpret voxels differently. This variability causes discrepancies in simulation results and the consequent predicted treatment outcomes. This study quantifies interobserver variability in breast MRI segmentation and its impact on MRgFUS simulation results. Basic segmentation training—including anatomy, MRI, and software instruction—was provided to ten research assistants with no prior segmentation experience. Each research assistant segmented the same breast MRI dataset into tissue types using the software Seg3D. To quantify interobserver variability, segmentations were compared using tissue volume differences and overlap-based metrics. Acoustic and thermal simulations were performed using each segmentation. The resultant thermal dose volumes and maximum temperatures were compared across segmentations to quantitatively assess the effect of interobserver variability. Overlap metrics indicate that the large background has the least interobserver variability while the small biopsy marker has the most variability. Acoustic and thermal simulations indicate that interobserver variability introduces significant variation in both thermal dose volume, varying from 0.1100 cm3 to 0.3409 cm3, and maximum temperature rise, varying from 36.89 °C to 70.86 °C for a single 30-second ultrasound sonication and 10 second cooldown. Comparing overlap-based metrics with tissue volumes can help contextualize which tissue type is most influenced by interobserver variability. The large variations in thermal dose and maximum temperature rise from thermal simulations indicate that interobserver variability greatly affects treatment planning. Future work will utilize these results to identify acceptable variation in labelled training data to use in fully automated machine-learning segmentation algorithms.