Bridging the Markerless Tracking Validation Gap: Intraoperative X-ray Workflow Integration

1 Introduction and Motivation

• Existing clinical markerless tracking data relies on surface landmarks and topology, not subsurface vertebral pose, while prior precise optical tracking required impractical exposure, hardware, or intraoperative CT.The paper identifies the absence of a clinically viable way to obtain precise vertebral-pose ground truth in real surgical settings.Source: §1 · p. 2
• The pipeline derives ground-truth vertebral pose in the RGB-D frame from routine intraoperative C-arm imaging.It avoids additional radiation, instrumentation in the surgical field, and deviation from standard workflows; ex vivo validation reported 2.06±0.29 mm and 3.35±0.93° accuracy.Source: §1 · p. 2
• The framework provides the first quantitative evaluation of markerless spine registration on clinical data against an independent subsurface reference.The authors report per-vertebra translational and rotational error across three surgical sequences and describe the framework as reusable benchmarking infrastructure.Source: §1 · p. 2
• The reported results are initial findings from a large-scale ongoing clinical data collection intended to support progressively larger benchmarking.The pipeline and evaluation framework are designed to scale as further cases are captured.Source: §1 · p. 2

2.1 Data Capture

• C-arm sources are modeled as pinhole cameras, with intrinsics calibrated using a multi-plane phantom and nine radio-opaque ArUco markers.Calibration images covered AP and oblique angles within ±30°; multi-view registration requires at least two fluoroscopic images from separate poses with known projection geometry.Source: §2.1 · p. 4

2.2 Intraoperative Capture

• A non-coplanar multimodal fiducial array is detectable in both X-ray and RGB-D images, linking coordinate frames without external tracking hardware.The array is positioned beside the exposure site, visible during routine vertebral-level checks, and covered to maintain the sterile field.Source: §2.2 · p. 4

2.3 CT Imaging to RGB-D Registration

• For each C-arm image, ArUco detections and PnP recover the board-to-detector transform, and shared marker coordinates provide relative transforms between views.The resulting projection matrices support intensity-based 2D–3D registration that optimizes six rigid-body parameters by maximizing mutual information between DRRs and fluoroscopic images.Source: §2.3 · pp. 4–5
• The registration process is repeated for each segmented vertebra to obtain its transform to the reference detector frame.The described transform chain connects the vertebral CT pose to the RGB-D camera frame.Source: §2.3 · p. 5

B and TRGBD

• ArUco detection and PnP pose estimation recover the marker array transforms in the reference detector and RGB-D camera frames.These transforms provide the frame-link information used in the vertebra-to-camera chain.Source: B and TRGBD · p. 5
• The complete transform chain maps the vertebra coordinate system to the camera coordinate system.The chain is presented as the means of locating vertebral pose in camera coordinates.Source: B and TRGBD · p. 5
• The displayed transform expression includes a product of coordinate transforms in the vertebra-to-camera chain.The expression is associated with the complete chain described in the section.Source: B and TRGBD · p. 5

3.1 Ex Vivo Validation

• Validation compared fluoroscopic and RGB-D marker-array tracking with optical tracking, then compared pipeline-derived vertebral poses with directly tracked poses on four excised ovine vertebrae.The end-to-end test used two C-arm shots with 30° angular disparity, comparable to the 25.6° mean disparity in captured clinical cases.Source: §3.1 · pp. 5–6
• 2.06±0.29 mm combined TRE and 3.35±0.93° rotational error: ex vivo validation against fiducial-tracked ovine vertebrae establishes the pipeline’s end-to-end accuracy.The authors report that this precision is sufficient as independent ground truth for evaluating markerless methods with substantially larger errors.Source: §3.1 · p. 6, Table 2 · p. 6

3.2 Clinical Data Registration

• Clinical data encompassed 11 vertebrae across 1,500 frames; the study compared the kinematic-chain strategy and fully automated per-vertebra registration, evaluating both against fluoroscopic pose estimates at a 3 mm correspondence distance.Source: §3.2 · p. 7
• 4.52°±1.00° vs. 8.36°±3.69°: the rigid method of outperforms per-vertebra ICP on rotation, while translation errors are comparable and both methods exceed cited navigation thresholds.Both methods have mean translational errors below 5 mm, but neither meets the ≤2 mm and ≤5° thresholds.Source: §3.2 · pp. 7–8, Table 3 · p. 8

• The clinical comparison enables evaluation of competing markerless methods against an independent subsurface reference, beyond prior clinical evaluations based on surrogate surface metrics.The pipeline derives per-vertebra 6-DoF pose from routine fluoroscopy without additional hardware or workflow disruption.Source: §4 · p. 8
• The rigid method of is meaningfully more accurate in rotation: its 3.84° advantage exceeds the 3.35±0.93° reference uncertainty, while translation errors are not statistically different.Registration errors measured against fluoroscopic estimates are upper bounds on true registration error; manual deformation tuning remains a deployment barrier.Source: §4 · p. 8
• Clinical evidence is limited to three sequences, with additional labeled cases from the ongoing study needed to broaden evaluation; 2D–3D registration accuracy depends on image quality, angular separation, and fluoroscopic view count.Source: §4 · p. 8
• The exposed posterior vertebral surface is geometrically ambiguous, limiting ICP’s ability to resolve pose, particularly rotation; a learned mapping could use appearance, anatomical context, and priors beyond ICP’s scope.Source: §4 · p. 8