Module
A contour that looks plausible and is wrong is the expensive failure
Contour review is hard because a bad contour rarely looks bad. It looks like a contour. This module makes the uncertainty spatial, compares against what was approved before, and quantifies what the difference would cost in dose.
A single confidence number hides the thing you need to see
Telling a clinician that a structure is 91% confident is close to useless: it does not say where. A structure can agree well overall and be badly wrong at the one boundary that sits inside the high-dose region. The useful output is a map of where the draft is least certain, a comparison against the previously approved version, and a statement of what the difference changes downstream.
What the module does
- Names structures correctly
- Resolves draft names against your institutional schema and standard nomenclature, so every downstream rule can parse them without a mapping step.
- Shows uncertainty spatially
- A heat map over the image, not a single number. Where the draft is least certain is where a reviewer should look first.
- Runs geometric checks
- Discontinuity, islands, overlap, containment, abrupt termination, laterality, volume, surface irregularity and body-boundary violation.
- Compares against the record
- Dice, surface Dice, 95th-percentile Hausdorff distance, mean surface distance and volume difference against the prior approved structure and against reference cohorts.
- Quantifies dose impact
- What changes in the reported metrics if the reviewer accepts the draft rather than the prior version.
- Captures every edit
- Each correction is recorded with the region, the reason category, a free-text rationale, the reviewer's role, the time spent, whether it was clinically material and what it changed downstream.
What it does not do
Boundaries of this module
- Approve contours. Final contour approval is a human gate.
- Define targets autonomously. Target definition is a higher-risk function than organ-at-risk drafting and is treated separately.
- Retrain on your corrections. Corrections become candidate training data behind adjudication and an offline release.
- Run outside its validated distribution. A scanner or acquisition outside the validated set is a stop condition, not a degraded result.
See it working
On the synthetic thorax case, one structure is discontinuous across four slices inside the high-dose region. The agent that drew it reports the gap as a correct anatomical boundary at 0.74 confidence. The independent checker disagrees and blocks. Both positions are shown side by side, and a person decides.
Status