Skip to content
Saif Khan
Chapter 03.02 · Collaborate

Spine Vision Research

Spine pose estimation, spine motion analysis, and AI for spine biomechanics across radiographs, CT, MRI, structured light, RGB, and RGB-D.

Work on spine vision is scattered across medical imaging, posture screening, surgical navigation, and markerless motion analysis. I made this page to put those threads next to each other: what people measure, what datasets exist, and which questions are still hard to test.

standing radiographs and biplanar X-ray CT and MRI anatomy modeling surface topography and structured light RGB and RGB-D posture, motion, and surgery benchmarking, inference, and biomechanics
SIMSPINE teaser showing multi-view human motion with spine-aware annotations
SIMSPINE represents one RGB-driven branch of a much broader spine-vision landscape.
This page covers
  • clinical imaging and deformity analysis
  • surface posture systems and structured light
  • markerless motion, benchmarking, and inference
6
sensor families

from clinical imaging to markerless sensing

8
representative resource lanes

benchmarks, systems, and datasets compared below

365
studies in a 2023 scoping review

showing how large the literature already is

15%
externally validated

reported by the scoping review of deep learning for spine care

Mission

Why this page exists

Public discussion of spine pose estimation, spine motion analysis, and spine biomechanics AI is still scattered: radiology papers focus on anatomy and pathology, commercial posture systems focus on repeated screening, and motion-focused computer-vision work rarely connects back to vertebral biomechanics. This hub compresses that landscape into one map.

  • The field is strongest on static anatomical understanding in CT and MRI, where shared segmentation benchmarks already exist.
  • Radiograph and biplanar systems are clinically important for deformity and alignment, but open ML workflows are less standardized.
  • Surface topography and structured light enable repeated, radiation-free follow-up, yet public benchmarking remains sparse.
  • Markerless RGB and RGB-D systems are only beginning to expose spine-aware motion, vertebral kinematics, and deployable inference.
Sensor atlas

Modalities and systems

These cards separate sensing modalities by what they measure well, how mature their evaluation is, and where public reproducibility still falls short.

Clinically established

Standing radiographs and biplanar X-ray

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.

vertebral landmarks and corners Cobb angle and global alignment patient-specific 3D reconstruction from paired views
  • single-view projection ambiguity
  • benchmarking is narrower than clinical usage
  • commercial 3D workflows are more mature than open ML tooling
Strong public benchmark

CT anatomy and vertebra segmentation

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.

vertebra localization and identification instance segmentation morphology and fracture-oriented measurements
  • structure is captured better than motion
  • benchmarks emphasize anatomy, not downstream biomechanics
  • cross-scanner robustness is improving but still matters
Growing benchmark ecosystem

MRI and soft-tissue segmentation

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.

vertebra and disc segmentation canal and foraminal structure delineation degeneration and pathology assessment
  • protocol heterogeneity across scanners and institutions
  • challenge infrastructure exists, but fewer dominant public baselines
  • biomechanics interpretation is mostly indirect
Clinically useful, weakly benchmarked

Surface topography and structured light

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.

back-surface shape and asymmetry postural screening and progression tracking non-radiographic follow-up measurements
  • surface geometry is only an indirect proxy for vertebral anatomy
  • public datasets are limited
  • commercial devices dominate over open research infrastructure
Emerging research lane

RGB and multiview markerless motion

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.

2D spine-centric landmarks multiview 3D spinal keypoints vertebral rotations and motion curves
  • detailed ground truth is scarce
  • annotation policies differ sharply between datasets
  • benchmarking is only now becoming reusable
Early but strategically important

RGB-D and intraoperative sensing

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.

depth completion and point clouds surface-to-anatomy registration guidance-ready geometry for navigation
  • datasets are smaller and more specialized
  • few field-standard benchmarks exist yet
  • generalization outside operating-room conditions is unresolved
Task stack

What spine vision is actually trying to predict

The field is broader than segmentation. These task families clarify why some modalities look mature and others still look structurally thin.

Localization and labeling

Identify vertebral levels, centroids, landmarks, discs, or anatomical key points before downstream measurements.

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.

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.

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.

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.

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.

Promising but specialized, with fewer shared public benchmarks than diagnosis-oriented tasks.
Comparison matrix

Benchmarks, systems, and representative releases

This matrix is intentionally cross-modal. It mixes public benchmarks, commercial clinical systems, and open research releases so visitors can compare what is measurable, what is reproducible, and what is still missing.

Resource Modality Primary focus Supervision Evaluation Access and tooling Biomechanics value
Challenge benchmark · Established but narrow
AASCE / SpineWeb
radiographs landmarks Cobb angle
Single-view radiographs Scoliosis landmark detection and curvature estimation Annotated vertebral landmarks and alignment targets on AP X-rays Challenge-style ranking centered on landmark and curvature accuracy
Public challenge resources via SpineWeb and Grand Challenge
Benchmark-oriented ecosystem, but no dominant reusable inference stack
Useful for coronal alignment and deformity; not dynamic
Clinical system · Deployed clinically
EOS / sterEOS
biplanar 3D reconstruction weight-bearing
Biplanar low-dose X-ray Weight-bearing 2D/3D reconstruction and alignment analysis Expert-guided reconstruction from paired standing radiographs Clinical workflow rather than open leaderboard
Commercial platform
No public inference stack
Very high for structural alignment under load
Public challenge + dataset · Strongest open anatomy benchmark
VerSe
CT segmentation labeling
CT Vertebra localization, segmentation, and identification Volumetric vertebra masks and labels across heterogeneous CT scans Standardized Grand Challenge tasks with shared metrics and reports
Public benchmark
Open challenge reports and baseline implementations are available
Indirect; excellent for anatomy, weak for motion
Public challenge + dataset · Growing open MRI benchmark
SPIDER
MRI segmentation lumbar
Lumbar MRI Lumbar spine segmentation and clinically relevant structure delineation Pixelwise annotations for lumbar structures on MRI Continuous challenge setup with hidden test evaluation
Public challenge infrastructure
Shared baselines and evaluation framework
Indirect; more anatomy and pathology than motion
Clinical system · Useful in practice, weakly benchmarked
DIERS formetric 4D
surface topography screening structured light
Structured light / rasterstereography Back-surface posture analysis and radiation-free monitoring Surface-shape capture and anatomical proxies rather than direct vertebral labels Clinical repeatability and correlation studies
Commercial system
Public code and shared ML benchmarks are sparse
Moderate; useful for posture and follow-up, indirect for vertebral anatomy
Research dataset / method · Emerging lane
SpineDepth
RGB-D surgery registration
RGB-D and point clouds Dense 3D reconstruction of exposed spine anatomy for surgical registration Paired RGB-D depth completion and anatomy-aware geometry targets Research-task metrics for depth and geometry quality
Research release
Useful reference point, but not yet a field-standard benchmark stack
High anatomical fidelity in procedure-specific settings
Paper + dataset · Available now
SpinePose
RGB 2D pose inference
Monocular RGB 2D spine-centric landmark estimation in natural images Manual 2D spine landmarks with biomechanics-aware evaluation Public paper benchmarks on spine-centric 2D tasks
Paper page, dataset, and inference entry point
Inference library available
Moderate; stronger than standard pose skeletons, still 2D
Paper + dataset + benchmark · Available now
SIMSPINE
RGB 3D motion benchmark
Multiview RGB with simulation-derived supervision 3D spinal keypoints, rotational kinematics, and benchmark baselines Simulated vertebral markers, multi-view detections, and inverse-kinematics labels Reference baselines across 2D, multiview 3D, monocular 3D, and biomechanics-aware checks
Paper page plus dataset on Hugging Face
Benchmark assets released; inference support is the next tooling step
High; anatomy-aware motion is the core point
My contribution in context

Where my own research sits

My work currently lives in the smallest but least-served part of this map: spine-aware understanding from external sensing, especially RGB and multiview motion, with a stronger emphasis on vertebral structure and biomechanics than conventional pose estimation.

Put differently: most of the open field is still strongest on anatomy under clinical imaging, while my current line is trying to make spine-aware structure and motion usable from external sensing.

Paper + dataset · Available now

SpinePose

Moves the field from generic human joints toward spine-centric 2D landmarks in natural RGB images.

Paper + dataset + benchmark · Available now

SIMSPINE

Extends the problem into multiview 3D supervision, vertebral kinematics, and biomechanics-aware benchmarking.

Open-source inference toolkit · Available now

SpinePose Inference Library

Turns released models into a usable CLI and Python API, so the work is not locked inside papers.

Adaptation benchmark · Adjacent research line

PoseAdapt

Targets continual adaptation and efficient optimization once pose models leave one fixed benchmark and meet changing domains.

Cross-cutting gaps

Where the field still looks structurally weak

These are the failure modes that show up repeatedly across modalities, even when individual methods look strong in isolation.

Static anatomy is easier to benchmark than motion

CT and MRI already support shared segmentation challenges, but vertebral motion, rotations, and full-body spine biomechanics still lack common public evaluation.

Clinical utility and open reproducibility diverge

Commercial systems like EOS and DIERS are important in practice, yet their maturity does not translate into open datasets, code, or challenge infrastructure.

Biomechanics validation is still exceptional

Many methods optimize localization or segmentation metrics without checking whether the predicted posture or motion remains anatomically plausible.

Cross-modal fusion is underdeveloped

The field still treats radiographs, MRI, surface scans, and RGB motion as separate silos instead of complementary evidence streams.

Deployment quality is lagging behind publication quality

Even when methods are strong on paper, reusable inference interfaces, benchmark harnesses, and failure-analysis tools are usually missing.

FAQ

Common questions about spine pose estimation and spine motion analysis

This section is meant to make the field easier to navigate for people arriving from search terms such as spine pose estimation, spinal imaging AI, markerless spine tracking, or AI for spine biomechanics.

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.

Selected sources

Representative papers, challenges, and systems

This first version is curated rather than systematic. It is meant to be genuinely useful now, then expanded over time.

Field review

Radiographs

Grand Challenge / SpineWeb

AASCE 2019 challenge

Representative public benchmark line for scoliosis landmarking and curvature estimation on AP radiographs.

Biplanar radiography

EOS imaging

EOS 3D modeling

Reference commercial workflow for weight-bearing 3D spine reconstruction.

CT

Grand Challenge

VerSe challenge

Key open benchmark for vertebra localization, labeling, and segmentation in CT.

MRI

Grand Challenge

SPIDER challenge

Representative open challenge for lumbar MRI segmentation and evaluation.

Surface topography

DIERS

DIERS formetric 4D

Commercial structured-light system for radiation-free posture and scoliosis follow-up.

RGB-D / surgery

RGB biomechanics

Project page

SpinePose

2D spine-centric RGB landmark estimation in unconstrained images.

Project page

SIMSPINE

Biomechanics-aware 3D supervision, multiview baselines, and vertebral kinematics from RGB.

Next layer

Help expand the map

Missing a public benchmark, a clinically important commercial system, or a spine-motion resource that should be here? Send it over. This page is designed to become a durable reference, not a static announcement.