Medical imaging annotation: X-ray & bone segmentation
Medical image annotation has a different center of gravity than any other domain: the boundary is the diagnosis. A fracture detector or joint-measurement model lives or dies on whether its training masks follow true anatomical boundaries-cortical edges, joint margins, fracture lines-rather than approximate blobs.
A dedicated medical segmentation engine
Projects created in Medical mode route to a segmentation model fine-tuned specifically on medical imagery-X-ray, dermoscopy, endoscopy, microscopy, ultrasound. It is concept-prompted: the label's clinical name ('femur', 'fracture line') is part of the prompt, so the model searches the image for that concept and the annotator's click simply disambiguates which instance is meant.
This matters because general-purpose segmentation models underperform on radiographs. They are trained on natural photos where objects have texture and contrast; bone edges are gradual intensity gradients with overlapping structures. A medically fine-tuned engine is the difference between a usable pre-label and a from-scratch hand-tracing job.
What we annotate in medical projects
- Bone & anatomy segmentation-per-bone instance masks on X-rays (femur, tibia, radius, vertebrae), joint regions, and implant hardware.
- Pathology localization-boxes or masks on fractures, lesions, opacities, and foreign objects.
- Keypoint annotation-anatomical landmarks for measurement models: joint centers, cortical reference points, orthopedic planning angles.
- Study-level classification-normal/abnormal tags, view labels (AP, lateral, oblique), and quality flags.
AI proposes, a human disposes
Model-generated masks are treated as fast drafts for expert correction, never final answers. The annotator selects the label, clicks the region, and the medical engine returns candidate regions; the click picks the intended one and the mask lands as an editable polygon, adjusted vertex-by-vertex wherever clinical judgment disagrees. Medical projects are expected to arrive de-identified, and datasets are only ever delivered back to the originating client.
Formats & delivery
All standard image formats are accepted (PNG, JPEG, frames exported from DICOM), annotated at native resolution so subtle fracture lines are not lost to downscaling. Exports ship as COCO JSON and PNG masks with versioned train/valid/test splits, and we can run double-annotation on a sampled subset to report inter-annotator agreement (IoU).