⚠ Local upload — this image is only in your browser, not stored. It may be PHI. Analyze with AI sends it to the consent-gated server pipeline; only do so for de-identified or consented images.
No box selected. Click a box, or switch to Draw to add one.
—
🔒 Never auto-applied. Decisions here are drafts saved in your browser + exportable.
The authoritative, verifier-gated write to the model registry happens in the
studio (local bridge on cortex-dev; apex studio-write API).
Samples are synthetic — no PHI on public paths.
Annotate & Approve — guide
Review AI-predicted findings on dental images, correct them, and record an approval decision. Everything is human-in-the-loop: nothing is applied to a model without a downstream verifier.
The three panels
Queue (left) — every image awaiting review. The colored dot shows its state: blue=pending, green=approved, yellow=changes, red=rejected. Click to open.
Stage (center) — the image with its finding boxes. Green boxes are human-edited; purple boxes are the model's original predictions.
Inspector (right) — edit the selected box's label & confidence, see all findings, write a reason, and record a decision.
Adding your own images
⬆ UploadAdd your own pictures. They're downscaled and kept local in your browser (never uploaded), enter the queue with an upload badge, and can be annotated + exported like any case. Any aspect ratio is handled.✨ Analyze with AISend the current image to the teacher-VLM server pipeline (/api/vision/upload → /api/vision/classify) for a classification prediction, added as a full-image box. Consent-gated (PHI) — needs the backend + a consent record configured; shows a clear error otherwise. A warning confirms before any send.
Editing boxes
Select modeClick a box to select it. Drag its body to move; drag the yellow corner handle to resize.Draw modeClick Draw box (or B), then drag on the image to add a new finding. It starts with the last-used label.RelabelWith a box selected, pick a new label from the dropdown. Edited boxes turn green.ConfidenceDrag the slider to record how sure you are (human boxes default to 100%).DeleteSelect a box and press Del or the trash button.ResetRestore this image's original model predictions, discarding your edits.
Recording a decision
✓ ApproveThe boxes are correct (or you fixed them). Marks the case approved and jumps to the next pending one.✎ ChangesMostly right but needs another pass — requires a reason. Sends it back as "changes requested".✕ RejectNot usable (wrong image, unusable quality) — requires a reason.
Saving & the authoritative write
Decisions are saved in your browser automatically and can be Exported as JSON. This command-center page is a read/review surface — it can't call the local control bridge (different origin), so the final verifier-gated write into the model registry is done in the Studio (Studio write →): local make cortex-dev bridge, or apex /api/cortex/studio-write.
Keyboard shortcuts
VSelect modeBDraw-box modeDelDelete selected boxAApprove & nextCRequest changesRReject← →Previous / next case?Toggle this help