dental-rad-cli

Chairside dental radiograph documentation aid. Takes bitewing (BW) or periapical (PA) X-rays, produces structured findings JSON, an annotated PNG, and a note-draft text — suitable for copy-paste into clinical notes. Two findings in v0: alveolar bone loss (severity + horizontal / vertical pattern) and interproximal caries (with depth staging).

This is a documentation aid, not a diagnostic device. Every finding is a candidate the doctor verifies before it enters a chart. No FDA path; no autonomous claims.

Status

Pre-v0 prototype. Hour-0 in progress: training pipeline cleanroom- reimplemented from arxiv 2506.20522 methodology, models training on RTX 4090 (hostname: pickles), first signal expected on a held-out set of scrubbed bitewings.

Design doc (decisions + rationale): ../work-root/output/proposals/2026-05-11-dental-rad-cli-v0-design.md

Scope

• In scope (v0): tooth detection + FDI numbering, CEJ / bone-crest / apex keypoints, tooth + bone segmentation, per-site % bone loss, horizontal / vertical pattern classification, AAP staging summary, interproximal caries detection (with depth staging if dataset supports), annotated PNG render, structured JSON output, template note-draft text.
• Out of scope (v0): periapical lesions, overhanging restorations, calculus, furcation, resorption, pulp findings, impactions, panoramic films, multi-image visit aggregation, web UI, Curve integration, per-doctor phrasing customization, FDA path.

Hardware

• Training: NVIDIA RTX 4090 (host pickles over Tailscale)
• Inference at chairside: any CUDA box or M-series Mac (MPS path TBD post-hour-5)

Datasets

• Bone loss: DenPAR v3 — 1,000 annotated intra-oral periapical radiographs, CC-BY 4.0, official Training / Validation / Testing splits
• Caries: Roboflow public BW datasets (specific dataset decided during hour-0 inspection)

Methodology reference

Cleanroom-reimplemented from arxiv 2506.20522 (Wimalasiri et al., Scientific Reports 2026). The upstream repo (chathurawimalasiri/analysis-in-detecting-alveolar-bone-loss) has no LICENSE file, so this project follows a methodology-reference discipline: read the upstream code to understand the approach; write fresh code in this repo; do not copy or port directly.

License

MIT. See LICENSE.

Layout

dental-rad-cli/
├── data/                  # gitignored — DenPAR + Roboflow datasets
├── weights/               # gitignored — trained checkpoints
├── docs/                  # PHI scrub recipe, hour-5 gate criteria
├── scripts/               # data download + training entrypoints
├── src/dental_rad_cli/    # pipeline modules, CLI entrypoint
├── tests/                 # rule-layer unit tests
└── examples/
    ├── input/             # synthetic samples for smoke tests
    └── eval/              # 3-5 scrubbed BWs/PAs for hour-0 evaluation

Run (post-hour-0)

uv sync
bash scripts/download_denpar.sh
bash scripts/train_all.sh         # ~3-5 hours on RTX 4090
uv run dental-rad-cli analyze examples/eval/bw01.jpg --out ./results/

Cold-pickup install < 30 min once weights/ is populated.

Eval (held-out test sets)

bash scripts/eval.sh

Headline metrics:

• cej_collapse_rate — CEJ keypoint head, DenPAR Testing split, 200 PAs. Fraction of high-confidence predictions where mesial/distal CEJ < 10 px apart (lower is better).
• caries_map50 / caries_map50_95 — caries head, Baasils ICCMS test split.

Baseline at HEAD: cej_collapse_rate = 0.3071, caries_map50 = 0.6478 (val, n=58).

Autoresearch

Karpathy-style overnight research loop targeting the CEJ collapse metric: see autoresearch/cej-collapse/program.md. One file the agent edits, fixed wall-clock budget, results logged to results.tsv.

Why we evaluate

See output/2026-05-12-lessons-learned-26hr-session.md for the pattern catalog of what burned 26 hours of human time before either metric existed.