Healthcare AI · Product safety
Turning visual AI output into a reviewable clinical workflow
A completed visual AI work package inside an ongoing R&D program—designed around data quality, uncertainty, clinician review and explicit release gates.
This case study is anonymised and focused on product decisions. It contains no patient, clinic or company-sensitive data.Data contract
Public images, annotations and quality report
Model evidence
Class-level performance, not one average
Human loop
Accept · Correct · Reject · No suggestion
Release gate
Feature flag, audit trail and silent validation
Evaluation snapshot
Public data · train 9,481 / validation 2,871 / test 1,580 · not clinical validationPublic dataset source ↗Context
The work package explored finding suggestions on panoramic dental images. The product problem was not simply detection accuracy: it was how to make uncertain model output inspectable and correctable without presenting it as autonomous diagnosis.
My role
I directed the product scope, clinician-review workflow, confidence and no-reliable-suggestion states, feedback loop, integration framing and release gates. Model and engineering work was delivered with the technical team; I do not present it as solo ML engineering.
The difficult decision
The public research dataset contained 13,932 images and 136,892 bounding-box annotations, plus invalid and out-of-bounds boxes that required an explicit quality contract. Performance varied sharply by finding class, so one global confidence story would have been misleading.
Approach
We framed the surface as clinician-reviewed suggestions behind a feature flag: accept, correct or reject. Low-confidence output can return no reliable suggestion. Class-level performance, auditability and future silent-mode validation determine rollout—not a single headline metric.
Outcome and boundary
The visual work package and engineering handoff are complete for this R&D stage. Impacted-tooth detection reached 0.882 F1, while other classes remained lower. That evidence supports narrower routing and continued validation—not autonomous clinical use. Real-data retrospective review and prospective silent mode remain next-stage gates.
What I’d carry forward
A healthcare AI product earns trust by exposing its limits. The release decision needs class-level failure costs, reviewer correction data, drift monitoring and a safe state where the system declines to suggest.