SHIVA Spine Toolkit
The umbrella project. A trustworthy, fast, reproducible spine measurement and planning tool, built in the open, designed to compete with Surgimap on the measurements that actually matter — and to be auditable in ways Surgimap is not.
Trustworthy.
Fast.
Reproducible.
A spine clinic measures 30+ films a day. A surgeon stakes a 6-hour PSO on these numbers. Inter-rater agreement within ±2° is the table-stakes. Everything in the roadmap ladders to one of those three.
Math correctness, not feature count
If PI is wrong by 4°, you've planned the wrong surgery. Trust is built through math correctness, synthetic ground-truth tests, and inter-rater data.
Surgimap's killer feature is 90 s/case
Not the math — the speed. Magnifier loupe, sub-pixel snap-to-edge, keyboard nudges, endplate tool. Sub-90-second exit criterion.
Two raters, same image, ±2° agreement
This is a software problem (precision tooling) more than a clinical training problem. Address it in the UI and the math.
Do these in order.
Don't skip.
Working document for taking the prototype to a tool that earns its place in a deformity surgeon's workflow. Each phase has an explicit exit criterion.
Trust the math (2–3 weeks)
Fix PI normalization and PI–LL signs. Synthetic ground-truth tests. Audit log skeleton. Persistent disclaimer. Schema versioning. Exit: every measurement reproduces a known synthetic case within 0.5°.
Precision tooling (3–4 weeks)
Magnifier loupe, window/level, keyboard nudge, endplate tool. Exit: 13 landmarks in under 90 s. Two raters agree within 2° on every measurement.
DICOM (2 weeks)
In-browser DICOM parsing with dicom-parser. Auto-calibration from pixel spacing. De-identification mode. Exit: drag a real DICOM lumbar lateral, no calibration step, measurements correct.
Clinical analysis (2 weeks)
Schwab badges, Roussouly type, GAP score, age-stratified population norms. Exit: output is what a deformity surgeon reads, not raw angles.
Architecture refactor (4 weeks)
TypeScript with branded angle types. Vite + modular src/core, src/io, src/ui, src/ml. IndexedDB for image blobs. Schema migrations.
ML assist (6+ weeks)
ONNX-Web in-browser detector trained on VerSe / AASCE2019. Confidence-based flagging. Rule: the user always reviews and explicitly accepts.
Surgical planning
Virtual osteotomy. Target solver (given target PI−LL, solve required wedges). Achieved-vs-planned compare on post-op imaging.
Validation paper
30 films × 3 raters × 2 tools × 2 timepoints. ICC computation. Submit to Spine Deformity or European Spine Journal.
What NOT to do.
Don't add 3D — lateral films are 2D and EOS users don't need this. Don't add full-spine AI segmentation — surgeons want corners, not contours. Don't go Electron — single-file HTML is a feature. Don't build a server — no backend = no HIPAA scope. Don't pursue FDA prematurely — stay an engineering prototype, get the science right, license later. Don't add user accounts — export/import JSON is the sync model.
The Toolkit ships
as a constellation.
Each project below is independently usable, but they share the same audit, provenance, and reproducibility primitives.
Spinopelvic THA planning support →
The flagship. Surgeon-supervised, FHIR + DICOM SR exports.
Mechanical envelope for sagittal balance →
Geometry-driven, not atlas-driven. ΔLL→ΔSVA, PI–LL band, SVA targets.
Open longitudinal spine dataset framework →
Verified open-resource registry + clinical-FE longitudinal framework + synthetic SQLite cohort.
Energy-aware alignment skill →
Decision-support module with statistical analysis plan + risk register + synthetic validation dataset.
The MIT code is free. A clinical site license covers deployment support, integration help, hash-verified releases, and indemnification language. See commercial offerings →