Abstract / Summary
C-arm cone-beam computed tomography (CBCT) provides three-dimensional bony anatomy for spine navigation, whereas tracked ultrasound can provide local radiation-free anatomical information. Accurate fusion remains challenging because of ultrasound artifacts, incomplete surface visibility, and coordinate-transformation errors. This phantom study evaluated an ultrasound-C-arm CBCT fusion workflow focused on probe calibration, multimodal registration, and target localization rather than robotic puncture or needle insertion. The system integrated ultrasound, C-arm CBCT, optical tracking, robot-controlled probe acquisition, and a navigation workstation. An L1-L4 lumbar phantom was scanned in water. CBCT bone segmentation used a 400-HU threshold with manual refinement. Probe calibration used an optimized double-layer 4 N-wire phantom and internal held-out validation. Tracked ultrasound surfaces were registered to CBCT using similarity initialization, multiresolution point-to-point ICP, and TPS residual refinement. Ten independently acquired scans were evaluated at eight bilateral pedicle-center landmarks withheld from optimization; the run-level mean TRE was the statistical unit ( n = 10). Supplementary real-scan, retrospective, and computational analyses tested operator/pitch robustness, frame density, ICP formulation, and dynamic inconsistency. The ultrasound image scale was 0.16 mm/pixel. Full-data 4 N-wire calibration error was 0.627 mm (pooled 3D vector RMS calculated directly over the complete residual set); the corresponding per-image-set RMS across the 80 calibration image sets was 0.621 ± 0.086 mm. Split-half held-out error was 0.639 ± 0.052 mm. Point-to-point ICP TRE was 2.98 ± 0.23 mm and decreased to 1.26 ± 0.08 mm after TPS ( p < 0.001). Mean/RMS/HD95 whole-surface distances decreased from 1.80/2.05/3.61 to 0.77/0.88/1.54 mm. Point-to-plane ICP modestly improved the rigid stage but changed final TPS TRE by only − 0.005 mm. The effective ~ 170-mm acquisition region corresponded to ~ 34 s at 5 mm/s, and the cumulative machine-processing component time after scanning was 143.7 s; the TPS solve itself required ~ 0.06 s. The workflow achieved millimeter-level ultrasound-C-arm CBCT fusion accuracy in a controlled spine phantom. Further tissue-containing, cadaveric, animal, and clinical validation is required. The findings do not validate a complete robot-assisted surgical system because instrument execution was not assessed, dynamic analyses were computational, and the current human-in-the-loop workflow is not suitable for continuous real-time use.