The committed scope

Milestones

Weeks run sequentially but overlap intentionally — UI work begins against the pipeline as soon as it produces stable output, and packaging begins as the pipeline stabilises. If either detector needs more tuning time, periapical radiolucency is de-scoped first, not caries.

1. Week 1–2

   Foundation Layer (Module 0) — Stage 0 + Stage 1

   • Image quality gate (accept / enhance / retake)
   • ESRGAN enhancement, original always retained
   • Metal-artifact flagging
   • FDI tooth numbering — detect, segment, number
   • Persistence schema: tooth + date + model version
   • PHI redaction (Presidio + OCR)
2. Week 3–4

   Cariology + Endodontics core pipeline (v1.0 MVP)

   • Bootstrap from open pretrained weights (DENTEX, DentalXrayAI, DenPAR)
   • Caries detection across periapical, bitewing and panoramic
   • Enamel/dentin binary staging
   • Periapical radiolucency detection (Story 16)
   • Per-lesion mask and FDI attachment for both finding types
   • Confidence-band framework, tuned for high recall on periapical findings
3. Week 5

   Pearl-style output & clinician UI

   • Colour-coded lesion overlay for caries and periapical findings
   • Per-finding callout: class · stage · confidence
   • Eye toggle per finding type
   • Retake-prompt UI
   • Enhanced-vs-original toggle
   • Case summary view
4. Week 6

   On-device packaging & offline runtime

   • Model quantization (ONNX / TFLite)
   • Mid-range Android device target
   • Full offline inference pipeline
   • Latency budget: <5s per image, <1.5s combined for Stage 0+1
5. Week 7

   Local validation harness & Nigerian test set

   • Local annotation and validation tooling
   • Modality-stratified accuracy reporting scaffold
   • Sensitivity, specificity and false-negative rate, caries and periapical reported separately
   • No in-app accuracy claim until this gate is met
6. Week 8

   QA, hardening & demo handoff

   • End-to-end regression across three modalities and both finding types
   • Audit-trail verification on 100% of findings
   • Bug fixing and demo build
   • Handoff documentation and walkthrough

Assumptions and dependencies

Each of these is load-bearing. A3 in particular constrains what the product is allowed to say about itself for the whole of the sprint.

A1

Access to the open datasets referenced in the PRDs (DENTEX, DenPAR, DentalAI, Bangladesh OPG) is unrestricted for training use.

A2

A defined mid-range Android device or spec is agreed by end of Week 1 — it sets the latency and quantization targets.

A3

No in-app, marketing or regulatory accuracy claim is made at the end of this sprint. This sprint delivers the harness, not the claim.

A4

This module is decision support only — a second opinion for a licensed clinician, not autonomous diagnosis.

A5

The endodontic beachhead is limited to Story 16 only; Stories 17–25 are not assumed reachable within this sprint.