Skip to content
Saif Khan
Chapter 04.02 · Spine Vision

Sensing the Spine

No single sensor gives us anatomy, posture, and movement at once. These six approaches trade detail, repeatability, and freedom of movement in different ways.

Sensing approaches

Standing radiographs and biplanar X-ray

Clinically established. High value for deformity, alignment, and weight-bearing assessment.

Weight-bearing radiographs remain central to scoliosis, sagittal balance, and deformity analysis. Computer-vision work here focuses on landmarks, Cobb angle estimation, and 3D reconstruction from limited views.

Typical outputs: vertebral landmarks and corners; Cobb angle and global alignment; patient-specific 3D reconstruction from paired views.

Current limits: single-view projection ambiguity; benchmarking is narrower than clinical usage; commercial 3D workflows are more mature than open ML tooling.

Resources: AASCE challenge · EOS 3D modeling

CT anatomy and vertebra segmentation

Strong public benchmark. The best-covered open setting for vertebra labeling and segmentation.

CT provides rich anatomy, robust vertebral boundaries, and multi-center datasets that support automated labeling, segmentation, and morphometric analysis. It is the most benchmark-ready spine-imaging lane today.

Typical outputs: vertebra localization and identification; instance segmentation; morphology and fracture-oriented measurements.

Current limits: structure is captured better than motion; benchmarks emphasize anatomy, not downstream biomechanics; cross-scanner robustness is improving but still matters.

Resources: VerSe challenge

MRI and soft-tissue segmentation

Growing benchmark ecosystem. Strong for discs, vertebrae, canals, and degeneration-related structures.

MRI supports lumbar structure delineation without ionizing radiation and is central for degenerative disease, spinal canal analysis, and soft-tissue context. Open benchmarking exists, but is still less mature than CT.

Typical outputs: vertebra and disc segmentation; canal and foraminal structure delineation; degeneration and pathology assessment.

Current limits: protocol heterogeneity across scanners and institutions; challenge infrastructure exists, but fewer dominant public baselines; biomechanics interpretation is mostly indirect.

Resources: SPIDER challenge

Surface topography and structured light

Clinically useful, weakly benchmarked. Radiation-free follow-up and posture analysis are strong practical advantages.

Rasterstereography and structured-light systems measure back-surface geometry to infer posture and scoliosis-related asymmetries. They are attractive for repeated monitoring, but public ML datasets and common benchmarks are rare.

Typical outputs: back-surface shape and asymmetry; postural screening and progression tracking; non-radiographic follow-up measurements.

Current limits: surface geometry is only an indirect proxy for vertebral anatomy; public datasets are limited; commercial devices dominate over open research infrastructure.

Resources: DIERS formetric 4D · Surface topography review

RGB and multiview markerless motion

Emerging research lane. This is where unconstrained posture and motion finally meet biomechanics-aware supervision.

Conventional pose datasets barely represent the spine beyond shoulders and pelvis. Newer work starts to recover spine-centric 2D landmarks, multiview 3D positions, and vertebral rotations from natural full-body motion.

Typical outputs: 2D spine-centric landmarks; multiview 3D spinal keypoints; vertebral rotations and motion curves.

Current limits: detailed ground truth is scarce; annotation policies differ sharply between datasets; benchmarking is only now becoming reusable.

Resources: SpinePose · SIMSPINE

RGB-D and intraoperative sensing

Early but strategically important. Dense 3D sensing matters when anatomy must be registered during a procedure.

RGB-D and point-cloud methods are emerging for exposed-spine capture, navigation, and registration. The research emphasis is less on general posture and more on anatomically faithful geometry under procedure-time constraints.

Typical outputs: depth completion and point clouds; surface-to-anatomy registration; guidance-ready geometry for navigation.

Current limits: datasets are smaller and more specialized; few field-standard benchmarks exist yet; generalization outside operating-room conditions is unresolved.

Resources: SpineDepth

Next: Benchmarks and Systems →