Clinician-led quality validation for treatment-planning models
OVERVIEW
A dental/orthodontic company building CBCT-scan segmentation models to support clear-aligner and orthodontic treatment planning engaged us to build a clinical QA pipeline for validating its models' outputs.
For every patient case, three model outputs were validated, and each one went through consensus review by three expert orthodontists: crown segmentation, axes setup, CT segmentation (bone and teeth) on two modalities - X-Ray and MRI.
QUALITY
Three model outputs, one unified review team
The review team validated 3 distinct outputs from the client's pipeline: crown segmentation, axes setup (reviewed per jaw, upper and lower scored separately) and CT segmentation (bone and teeth annotations combined into a single review per case).
Consensus labeling with expert orthodontists
We ran a consensus-labeling process: three expert orthodontists independently reviewed the same task, and their judgments were reconciled into a single consensus label. This reduced the effect of individual clinical subjectivity on the delivered data and gave the client a defensible, multi-expert rationale wherever opinions diverged.
Structured anomaly taxonomy
Working with the client's QA and research team, our clinicians applied and refined a shared anomaly taxonomy covering missing and each tooth's identity, position, location, anomalies, etc. - standardized throughout on the FDI tooth-numbering system.
ITERATIVE PROCESS
Benchmarked against pre-reviewed cases
Each delivery was scored against the client's own pre-reviewed benchmark cases.
Clinician-to-clinician clarification loop
Our orthodontists raised edge-case questions directly to the client's clinical team and translated every answer back into updated internal review guidelines, keeping the two teams' clinical judgment aligned.
