Accuracy of inside-out visual-inertial tracking pipelines for freehand 3D ultrasound probe localization

Abstract

Freehand three-dimensional ultrasound often requires accurate probe localization. Inside-out visual-inertial tracking mounts the camera and inertial unit on the probe itself, offering a compact alternative to external systems. Achievable accuracy depends on the complete pipeline, not the sensor alone, and comparisons rarely state which differences their setup can resolve. We compared three pipelines under identical conditions: the ZED X Mini with the ZED SDK, the ZED X Mini with cuVSLAM, and the Intel RealSense D455 with cuVSLAM. Both cameras and an infrared-reflective marker cluster were mounted on a rigid frame representing the probe base, tracked by an eleven-camera motion-capture system. Four recordings covered three motion conditions of decreasing constraint, from pivoting about a fixed point to full freehand manipulation, one repeated to quantify trial-to-trial variability. All processing was offline. Trajectories were synchronized and aligned by robot-world/hand-eye calibration, and accuracy assessed from per-frame residuals. The ZED X Mini with cuVSLAM gave the lowest median rotational residual in every condition, the lowest median translational residual in most (2.5–4.1 mm and 0.100–0.117°), and the lowest 95th-percentile residual in every condition and component. Part of that advantage reflects cuVSLAM returning a full-sequence optimized trajectory, whereas the ZED SDK returns causal per-frame poses. Differences between motion conditions largely fell within the resolution of the comparison, derived from reference uncertainty and the precision of each statistic, the exception being a higher translational residual for ZED X Mini with cuVSLAM under freehand motion. Pipeline choice, more than motion condition, is therefore the dominant determinant of tracking accuracy, and the ZED X Mini with cuVSLAM is the most suitable of the three.

For freehand 3D-ultrasound applications, the rotational error is small relative to reported probe-to-image-plane calibration error, whereas the millimeter-level translational error dominates overall pose accuracy. Whether either is sufficient for volumetric reconstruction remains to be established.