AI jobs

The AI jobs page runs the full flow a dentist follows: pick a patient, pick the DICOM files (first checked is the primary document; check the rest of the series for volumetric nnU-Net runs), choose the backend (pano for single frames, nnU-Net for series volumes), and queue the run. Jobs list with their status — proposed, queued, running, done, failed, cancelled — plus errors and the runner log tail.

Opening the page

Open /imaging-ai and search the patient in the single patient control, or deep-link /imaging-ai?patient_id=<uuid> from the patient record — the control initializes from the URL. Without patient read rights, only the deep link works. The job list refreshes on its own on the scheduler-tick cadence; failed and cancelled jobs read "Not applicable" for review instead of "pending review".

Drafts and review

Finished runs ingest their overlays as AI draft documents — never gallery radiographs, never paired to the source. Review the draft overlay, then Confirm: proposed runs start executing, finished drafts are marked reviewed. Nothing the model produces enters the clinical record without that explicit step. AI output is a visualization aid only — never a diagnosis.