Before comparing methods, it helps to ask what each one is trying to measure. A vertebra mask from a CT scan and a motion curve from video answer different questions.
Task families
Localization and labeling
Identify vertebral levels, centroids, landmarks, discs, or anatomical key points before downstream measurements.
Where things stand: Strong in CT, improving in MRI and radiographs, still rare in unconstrained RGB.
Segmentation and structure extraction
Recover vertebrae, discs, canals, or back-surface geometry at dense spatial resolution.
Where things stand: Most benchmarked in CT and MRI, with commercial strength in surface topography.
Alignment and deformity quantification
Estimate Cobb angle, sagittal balance, curvature, and postural asymmetry for screening or planning.
Where things stand: Clinically central in radiographs and topography, but cross-system evaluation remains fragmented.
3D reconstruction under load
Infer patient-specific 3D spine structure under standing, weight-bearing conditions.
Where things stand: Clinically mature in EOS-style workflows, but not broadly open in ML form.
Dynamic motion and biomechanics
Move beyond static posture into vertebral motion, rotational kinematics, and anatomically plausible movement.
Where things stand: Still the most under-benchmarked lane; this is where RGB and simulation are starting to matter.
Intervention support and registration
Use imaging or RGB-D geometry to support navigation, registration, or procedure-time guidance.
Where things stand: Promising but specialized, with fewer shared public benchmarks than diagnosis-oriented tasks.
Useful distinctions
What is spine pose estimation?
Spine pose estimation refers to predicting spine-related landmarks, curves, segments, or vertebral structures from images or video. Depending on the sensing setup, the target may be 2D spinal keypoints in RGB images, vertebra segmentation in CT or MRI, or posture and curvature estimates from radiographs or surface topography.
What is the difference between spine pose estimation and spine motion analysis?
Spine pose estimation usually refers to recovering structure or posture from a single image or frame, while spine motion analysis focuses on how the spine moves over time. Motion analysis therefore needs temporal consistency, 3D geometry, and ideally biomechanics-aware outputs such as vertebral rotations, curvature changes, or anatomically plausible segment motion.
Which sensing modalities are used for AI in spine biomechanics?
The field spans standing radiographs, biplanar X-ray, CT, MRI, structured light and rasterstereography, RGB and multiview video, and RGB-D or point-cloud sensing for surgical or navigation settings. Each modality exposes different tradeoffs between anatomy, motion, radiation, repeatability, and deployment cost.
Where can I find open datasets and benchmarks for spine motion and spine imaging AI?
Open resources exist, but they are unevenly distributed across the field. VerSe is a strong open CT benchmark, SPIDER represents lumbar MRI segmentation, AASCE covers scoliosis landmarking on radiographs, and my own SpinePose and SIMSPINE pages cover open RGB and biomechanics-aware motion resources.
Next: Sensing the Spine →