Quality pre-check
Exposure, contrast and sharpness measured before anything is sent for analysis — with a specific reason, never a bare retake, and never a block.
Stage 0 model committed; a browser heuristic stands in today.
Chairside Diagnostic AI
A second opinion, anchored to the tooth
Chanieldx numbers every tooth on a radiograph, flags what it sees, and hands the decision straight back to the clinician. Built for the chair, for modest hardware, and for connectivity that comes and goes.
Decision support — a second opinion for a licensed clinician, not autonomous diagnosis.
- 0.904
- mAP50, FDI tooth numbering Built and verified. The anchor every finding hangs off.
- 3
- Modalities in Cariology scope Bitewing, periapical and panoramic.
- 100%
- Findings carrying full provenance Tooth, capture date and model version on every row.
- 0
- Accuracy claims made None until the Nigerian validation set is scored. By design.
How it works
Four things happen to every capture
One continuous flow from the chair, with a place for the clinician to intervene at every step.
Tooth numbering
Every detected tooth gets its FDI number. Nothing is charted against a raw pixel coordinate, and a wrong number is one tap to correct.
Built and verified — mAP50 0.904.
Findings overlay
Colour-coded lesion overlays with per-layer toggles and a callout for every finding. Deliberately not red/green — that would read as a verdict.
Built on the live analysis service.
Review and record
Accept, correct or dismiss. A dismissal stays distinguishable from the model finding nothing, and every decision is attributable.
Built, with a full audit trail.
Colour-coded, and deliberately not red or green
A red/green overlay reads as a pass/fail clinical verdict. This product does not make one. Each finding type gets a distinct hue chosen for differentiation, with related findings sharing a family and out-of-scope classes deliberately muted.
Every layer can be switched off independently, so a cluttered multi-finding image simplifies on demand rather than forcing the clinician to read through it.
- Caries Cariology mAP50 0.35
Real but weak — trained on bootstrap data only. Expected to improve once proprietary Nigerian radiographs are annotated.
- Deep Caries Cariology mAP50 0.50
A coarse binary depth signal, not the full E1–D3 staging the 53-story vision calls for.
- Periapical Lesion Endodontics mAP50 0.31
This is DENTEX's panoramic-modality class (Story 17, deferred) — not the Story 16 intra-oral periapical model, which has no trained weights yet. Nine validation boxes; not a trustworthy number in either direction.
- Impacted Tooth OMFS (incidental) mAP50 0.91
Comes along for free from the shared dataset. Strong result, but outside this project's committed clinical scope — presented muted and unemphasised.
The posture
Decision support, stated plainly
The constraints below are design decisions taken from the specification, not disclaimers added at the end.
Never autonomous
Every output is a suggestion for a licensed clinician to accept, edit or dismiss. That constraint shapes the screens rather than sitting under them.
No claim before the evidence
Confidence bands are designed and unfilled. Thresholds need the local validation harness, and inventing them now would produce a number that looks validated and is not.
Nothing detected is not healthy
An empty result is worded as an absence of detection, never as clinical clearance. That copy is fixed in one place and reused everywhere.
Honest about what is unbuilt
Every capability carries a status. Sharing stays blocked until PHI redaction is real, and no screen implies a model that does not exist.
Built for the target environment
Older sensors, digitised film, unreliable connectivity
The quality gate is forgiving because the alternative reads as broken. Images are stored before inference is attempted, so a service outage never costs a capture. Existing cases stay fully reviewable offline.
Outage never costs a capture
Images are stored before inference is attempted. The chair has moved on by the time anyone notices the service was down.
No direct identifiers stored
Cases key on the pseudonymous reference the practice already assigns. No name, date of birth or contact detail.
Sharing stays blocked
Until PHI redaction is real and verified, nothing can be sent onward. A redaction-uncertain state must block, not warn and proceed.
Working today
Not a roadmap. These are the capabilities that respond right now, tested against the live deployment.
FDI tooth numbering (Stage 1)
mAP50 0.904. The anchor every finding attaches to.
POST /api/v1/analyze
One image in, one tooth-anchored findings response out. No auth, no persistence, no banding.
Persistence (tooth + date + model version)
Implemented in this platform: every finding carries its provenance triplet and a full audit trail.
Findings overlay UI
Implemented in this platform: colour-coded overlays, per-layer toggles, per-finding callouts, review controls.
Authentication & practice accounts
Implemented in this platform: sessions, roles and per-practice scoping.
Start with a case
Open a visit, capture a radiograph, and see the numbered teeth and findings come back attached to it.
Chanieldx is decision support. It is not a medical device clearance and not for autonomous diagnosis.