Evaluating Optimization-Based Image Refinement for 3-D Ultrasound Mosaicking: A Sensitivity Analysis of Clinical and Benchmark Data

This study evaluates the performance and sensitivity of an optimization-based registration framework for mosaicking partially overlapping 3-D ultrasound volumes to expand the clinical field of view. We specifically characterize the algorithm's ability to recover from varying degrees of initial tracking uncertainty across different anatomical contexts.

Performance was assessed on (i) a synthetic benchmark knee data set (240 volume pairs from 16 volunteers) where controlled perturbations (±3 mm/°) were applied to a known ground-truth pose to evaluate convergence limits, and (ii) a clinical abdominal data set (128 volume pairs from 6 volunteers, 3-5 partially overlapping volumes per session), where the algorithm was initialized both at the reference optical tracking pose and under broader simulated tracking errors (up to ±10 mm/°).

On the synthetic data set, the method achieved high registration success (95.4%-96.2%) and accuracy (93.3%-93.8%). On the clinical data set, deviations from the reference increased with wider initial perturbations, highlighting a limited capture range. However, when initialized at the tracking-based position, the image-based refinement improved alignment compared to tracking alone in 44% of cases, with Dice gains up to 10%.

These findings demonstrate that while image-based refinement can enhance hardware-based mosaicking, its reliability depends heavily on the accuracy of the initial coarse alignment and the search constraints applied.