X-ray Bone Segmentation with a Dedicated Medical Engine
Pixel-accurate bone masks on X-ray imagery from our dedicated medical engine, where the label's clinical name is the model's concept prompt - human-verified, on isolated infrastructure.
The challenge
Medical segmentation punishes imprecision: a mask that bleeds a few pixels past the cortical boundary changes what a downstream model learns about bone geometry, and a landmark placed a few pixels off corrupts every angle computed from it. Generic segmentation models trained on natural images routinely miss those boundaries, because X-ray gradients behave nothing like photographs.
Clinical data also raises the bar on handling. Imagery cannot sit in shared annotation pools, reviewers need enough anatomical context to judge edge cases, and research teams need the annotation provenance documented well enough to survive peer review.
And the economics are brutal: the people qualified to segment medical imagery are expensive, so every minute spent hand-tracing a boundary that a model could have proposed is money burned.
Why this data matters
Annotation is widely acknowledged as the bottleneck in medical AI - models and compute are commoditizing faster than labeled clinical data. Public datasets cover a handful of anatomies and acquisition setups; anything else needs purpose-built annotation.
Bone segmentation on X-ray is deceptively hard: overlapping structures project onto the same pixels, exposure varies wildly between machines, and the clinically meaningful boundary (cortical edge, joint space) is often a subtle gradient rather than a crisp line.
The supervision type matters as much as the labels: detection boxes, segmentation masks, and landmark/measurement geometry train different kinds of models. Choosing the right one per use case - and being able to deliver all three - is half the value of a specialized medical annotation service.
How we did it
- 1
Medical engine
Segmentation runs on a dedicated engine tuned for medical imagery. The label's clinical name (femur, tibia, rib) is used as the model's concept prompt, so a click segments the named structure - not just a bright region. The standard engine covers non-clinical content.
- 2
AI proposes, experts verify
The engine returns an editable boundary in about a second; trained reviewers correct edge cases in a structured review queue. Every mask is human-approved before delivery - nothing ships on model confidence alone.
- 3
Landmarks & measurements
Where the use case is orthopedic geometry rather than region masks, anatomical landmarks and measurement constructs (axes, angles, distances) are annotated directly, so downstream models train on the geometry clinicians actually use.
- 4
Supervision matched to the model
Detection boxes for triage models, pixel masks for segmentation networks, landmarks for measurement models - the annotation plan is designed around what will be trained, not around what is easiest to draw.
- 5
Isolated handling
Imagery stays on dedicated private infrastructure - access-controlled storage, isolated processing, short-lived download links. We expect de-identified imagery, and data is never pooled across clients or used to train third-party models.
- 6
Versioned releases
Datasets ship as immutable versions with fixed splits and quality metrics, so a result in a paper or a regulatory filing can always cite the exact dataset version it was produced from.
Inside the dataset
- Pixel-accurate bone masks per structure, with clinical class names
- Anatomical landmarks and measurement geometry where required
- Region-level classification labels for triage-style tasks
- Per-label reviewer status - every mask carries its verification trail
- Acquisition-condition variety preserved (exposure, positioning, hardware)
Quality control
Medical review is structured in two tiers: the engine's proposal is corrected by a trained annotator, and ambiguous anatomy is escalated rather than guessed. Reviewer decisions are tracked per mask, which produces the audit trail research and regulatory workflows expect.
Quality metrics ship with each version: boundary-level agreement statistics on double-annotated samples, per-class review pass rates, and the escalation log. That is what turns a pile of masks into a publication-grade dataset.
What was delivered
- Pixel-accurate bone masks, human-verified
- Anatomical landmarks and measurement geometry where required
- Per-dataset quality metrics and reviewer audit trail
- Versioned exports suitable for pilot-to-publication workflows
Delivered as PNG class-index masks, COCO JSON with polygons or RLE, and landmark coordinate files, with fixed train/valid/test splits per version.
What the data enables
- Segmentation networks for bone and joint delineation
- Orthopedic measurement models (angles, axes, joint spaces)
- Triage and detection models over radiographs
- Multi-modal medical datasets when combined with reports or other imaging
Want a dataset like this?
Send us your images or video - we annotate, review, and deliver a training-ready dataset in your format.
Book an intro call