gregoryschwingmdphd/spinesurg-ct-annotator
Femoral head annotation
Two clicks per lateral radiograph: left femoral head centre, then right.
Why this exists
The hip point in XRSpinoPelvic1K is learned from synthetic DRRs, where the ground truth is a 3-D sphere fitted to the femoral head and projected. BUU — the only large public set of real lateral films — annotates L1–L5 and S1 and nothing pelvic, so there is currently no way to measure whether that synthetic hip point survives contact with a real radiograph. Without it, PI cannot be validated at all.
These annotations are a reference set, not training data. Test split only.
What you are marking
The hip axis is the line joining the centres of the two femoral heads; the point used for PI/PT/SS is its midpoint (Legaye & Duval-Beaupère, Eur Spine J 1998). The femoral head is very nearly a sphere, so its projection is a circle — you are marking the centre of that circle, a geometric centre rather than a surface point.
Method. Trace the subchondral cortical arc — the thin dense line of the articular surface — and mark its centre of curvature. This is the Mose concentric-circle template done by eye, which is why the worked example draws concentric rings. Use the 4× magnifier.
The trap. Do not aim at the brightest shadow. Overlap with the acetabulum and the opposite head puts the densest region medial to the true centre, so "centre of the bright blob" is biased rather than merely noisy.
Never centre on: the fovea capitis (the medial notch — a defect in the sphere), the greater trochanter, the femoral neck or head–neck junction, the acetabular roof or teardrop.
One circle or two. On a well-positioned lateral the heads superimpose — mark it as LEFT and leave RIGHT empty. If rotation separates them, mark both; the midpoint is derived and the separation is recorded, because a wide separation means an oblique film and the parameters from it are less trustworthy.
Skip for a prosthesis, heads outside the collimated field, or an exposure where the cortical arc cannot be traced.
The example is a synthetic radiograph on purpose: there the centre is a 3-D sphere fit projected through the imaging geometry, so it is objectively correct rather than one annotator's opinion. A marked-up real film would put whoever marked it into every reader's head.
Rules that make the reference worth having
No automatic proposal is ever shown. Not the model's prediction, not the classical circle fit. Displaying a starting point would anchor the annotator to it, and agreement would then measure suggestibility rather than anatomy.
Both heads, separately. The bicoxofemoral point is defined as the midpoint of the two femoral head centres (Legaye & Duval-Beaupère 1998). Marking both lets the midpoint be derived, lets their separation flag an oblique film, and lets you mark one when the other is genuinely invisible.
Skip freely. A forced guess on an unreadable film is worse than no annotation — this set exists to be the reference, so a fabricated point becomes a fabricated error in whatever it validates.
Two annotators per film, with the mean taken on agreement (≤2% of image width) and disagreements escalated to an adjudicator. Same claim/slot/TTL model as the segmentation review Space, so it should feel familiar.
Setup
Space secrets:
HFTOKEN write token for the ledger dataset ANNOTREPO <org>/xrsp-femhead-annot IMAGE_REPO <org>/xrsp-femhead-images ADJUDICATORS comma/space-separated HF usernames
Seed the ledger and upload the films:
python annot/seedcases.py --buu data/BUU-LSPINE --splits data/buusplits.json \ --annot-repo <org>/xrsp-femhead-annot --image-repo <org>/xrsp-femhead-images \ --n 120 --apply
Annotators paste their own HF token; the Space verifies it with whoami and uses the returned username as identity. It caches only sha256(token) -> username, never the token. Films are streamed through the Space, so annotators need no read access to any dataset.
For readers — how to start
- Make a HuggingFace account at <https://huggingface.co/join> if you do not have one (about a minute, free). Tell Greg your username so your reads are attributed.
- Open the annotator link and press Sign in with HuggingFace.
- Read the Criteria panel once, then collapse it. It carries the definition, the reference figure and a worked example.
- For each film: click the left femoral head centre, then the right one if rotation has separated them, and press Enter.
The controls
The magnifier follows your cursor at 4×. The next film is fetched while you work on the current one, so Enter should be instant.
"Femoral head not visible" is an answer, not a failure
Press it for a prosthesis, heads outside the collimated field, or an exposure where the subchondral arc simply cannot be traced. It counts as your read and still needs a second reader to agree — the fraction of films with no placeable hip point is one of the things this study is measuring, so recording it honestly is the job. Please do not guess: a fabricated centre becomes a fabricated error in everything measured from this set.
Use Pass only to hand a film to someone else.
Two reads per film
Every film is read independently by two people, and nobody is ever offered a film they have already read. If the two reads agree within tolerance the mean is taken; if they disagree, or if one reader marks a point the other could not see, the film is held for adjudication rather than averaged.
Progress
/board shows films finalised, reads done against the two-per-film target, how many await a second read, how many need adjudication, the not-visible rate, and a per-reader table. It refreshes every 10 seconds.