Inter-observer variability in breast segmentation and its impact on focused ultrasound thermal therapy modeling

Objective. Simulation-based treatment planning (SBTP) could improve magnetic resonance-guided focused ultrasound (MRgFUS) breast cancer treatments. Medical image segmentation is an essential SBTP step and may introduce uncertainty through interobserver variability. This exploratory study quantifies interobserver variability in breast magnetic resonance imaging (MRI) segmentation and assesses its impact on simulated MRgFUS breast cancer treatment outcomes.Approach.Nine researchers segmented three breast MRI. Interobserver variability amongst segmentation datasets was quantified using the surface Dice coefficient (sDICE), modified Hausdorff distance (MHD), and a multi-label, volume-weighted Dice similarity coefficient (wDICE). Segmentations were used in acoustic and thermal simulations to evaluate MRgFUS treatment outcomes including maximum temperature rise, thermal dose volume (TDV), and distance between the target location and simulated thermal center of mass (TCOM). Exploration of which labels and metrics best explained variation in treatment outcomes was performed with multiple regression using the ordinary least squares method and leave-one-out cross validation.Main results.Interobserver variability led to substantial differences in simulated outcomes, with maximum temperature changes up to 62.9 °C, TDV differences up to 324 mm³, and target-to-TCOM shifts as large as 3.6 mm. Regression analysis identified wDICE as a stable global metric across outcomes, while sDICE and MHD for the biopsy marker label highlighted label-specific variability. These results suggest wDICE as an overall indicator of segmentation variability for MRgFUS SBTP, with MHD and sDICE as potentially useful metrics for identifying labels where reduction of variability would be most impactful.Significance.Interobserver variability can meaningfully affect simulated MRgFUS treatment outcomes, warranting further study of clinical interobserver variability for SBTP. The exploratory regression modeling demonstrated here provides a potential methodology for identifying labels where reduction of interobserver variability would be most impactful.