OrthoAI v2 — AI-Assisted Clear Aligner Treatment Planning

OrthoAI v2 is a dual-agent deep learning pipeline for AI-assisted orthodontic treatment planning with clear aligners. It extends OrthoAI v1 by adding a CHaRM landmark-detection agent, a six-category biomechanical scoring model, and a multi-frame 4D staging simulator.

Paper: OrthoAI v2: From Single-Agent Segmentation to Dual-Agent Treatment Planning for Clear Aligners — Lansiaux E., Leman M., arXiv 2603.15663, March 2026.

---

What is and isn't in this repository

This project follows an open-core model :

Component	Status	License
DGCNN segmentation architecture	✅ Full source	MIT
CHaRM landmark detection architecture	✅ Full source	MIT
Abstract agent & orchestrator interfaces	✅ Full source	MIT
Synthetic benchmark + evaluation metrics	✅ Full source	MIT
FastAPI schema + route definitions	✅ Full source	MIT
React SaaS dashboard	✅ Full source	MIT
Evaluation scripts	✅ Full source	MIT
Manuscript (PDF + LaTeX)	✅ Released	CC BY 4.0
Pre-trained model weights	🔒 Not included	Commercial
Composite scoring calibration	🔒 Stub only	Commercial
Staging rule parameters	🔒 Stub only	Commercial
Training pipeline + data loaders	🔒 Not included	Commercial
Clinical movement limit database	🔒 Stub only	Commercial

Why this structure? The academic contribution (architectures, evaluation protocol, benchmark) is fully reproducible from this repository. The clinical calibration — the exact biomechanical constants, composite scoring weights, and staging parameters — constitutes a proprietary IP and is licensed separately.

This mirrors the model used by OpenAI (released GPT-2 architecture, not weights or RLHF pipeline), Mistral (released architecture + weights, not fine-tuning data), and MongoDB (open-source engine, closed Atlas cloud layer).

---

Quick start (open components)

git clone https://github.com/edlansiaux/orthoai-v2
cd orthoai-v2
pip install -e ".[dev]"

# Run evaluation on synthetic benchmark (200 crowding scenarios)
python scripts/evaluate.py --n_cases 200 --mode parallel

# Launch API with stub engine (demo data only, no real inference)
python -m api.main --stub

# Frontend dev server
cd frontend && npm install && npm run dev

Commercial engine (Orthalytix)

The production-calibrated engine (real biomechanical constants, trained weights, full composite scoring) is available via:

• SaaS API — api.orthalytix.com/v2 — contact us for access tokens
• On-premise license — Docker image with encrypted engine bundle
• Research license — free for academic institutions, contact below

---

Repository structure

orthoai-v2/
├── orthoai/
│   ├── models/
│   │   ├── dgcnn.py          # DGCNN architecture (MIT)
│   │   └── charm.py          # CHaRM architecture (MIT)
│   ├── agents/
│   │   ├── base.py           # Abstract agent interface (MIT)
│   │   ├── agent1_dgcnn.py   # DGCNN segmentation agent (MIT)
│   │   ├── agent2_charm.py   # CHaRM landmark agent (MIT)
│   │   └── orchestrator.py   # Fusion orchestrator (MIT)
│   ├── evaluation/
│   │   ├── metrics.py        # MEDE, MSR, quality score (MIT)
│   │   └── benchmark.py      # 200-case synthetic benchmark (MIT)
│   └── demo/
│       ├── synthetic.py      # Synthetic case generator (MIT)
│       └── presets.py        # 4 preset clinical archetypes (MIT)
├── engine_stub/
│   ├── __init__.py           # Public interface (MIT)
│   ├── scorer.py             # AbstractScorer + StubScorer (MIT)
│   └── README.md             # How to plug in commercial engine
├── api/
│   ├── main.py               # FastAPI app (MIT)
│   └── schemas.py            # Pydantic models (MIT)
├── frontend/                 # React dashboard (MIT)
├── scripts/
│   ├── evaluate.py           # Benchmark runner (MIT)
│   └── export_onnx.py        # ONNX export stub (MIT)
├── paper/                    # arXiv PDF + LaTeX source
├── docker/
│   ├── Dockerfile.open       # Open stack (no engine)
│   └── Dockerfile.commercial # Placeholder for licensed image
└── pyproject.toml

---

Citation

@article{lansiaux2026orthoaiv2,
  title   = {{OrthoAI} v2: From Single-Agent Segmentation to
             Dual-Agent Treatment Planning for Clear Aligners},
  author  = {Lansiaux, Edouard; Leman, Margaux},
  journal = {arXiv preprint arXiv:2603.15663},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.15663}
}

If you use the v1 baseline in comparison:

@article{lansiaux2026orthoaiv1,
  title   = {{OrthoAI}: A Neurosymbolic Framework for Evidence-Grounded Biomechanical Reasoning in Clear Aligner Orthodontics},
  author  = {Lansiaux, Edouard; Leman, Margaux; Ammi, Mehdi},
  journal = {arXiv preprint arXiv:2603.00124},
  year    = {2026}
}