ptv3-dental-landmark-detector

PTv3 Dental Landmark Detector

Automatic detection of anatomical landmarks on 3D intraoral scans, developed during a research internship at the National Dental Centre of Singapore (NDCS) and evaluated under the official 3DTeethLand — MICCAI 2024 challenge protocol.

The model takes a raw dental scan mesh (upper or lower jaw) and predicts six classes of clinically relevant landmarks — Mesial, Distal, Cusp, InnerPoint, OuterPoint, FacialPoint — directly from the point cloud, with no tooth segmentation stage.

Try the live demo Space →

Approach

Landmark detection is formulated as dense per-point distance regression:

1. Input — 10,000 vertices randomly sampled from the scan mesh, each with its surface normal → [N, 6] features (x, y, z, nx, ny, nz).
2. Encoder — a Point-Transformer-v3-inspired stack: linear embedding to 256-d, then 4 stages × 2 pre-norm transformer blocks (multi-head self-attention + MLP). Unlike standard PTv3, no downsampling is applied, so per-point resolution is preserved end to end.
3. Decoder — a lightweight per-point MLP (256→512→6) regressing, for each landmark class, the normalized distance from the point to the nearest landmark of that class. Training targets are Euclidean distances clamped at 15 mm and sqrt-sharpened: d = sqrt(min(d, 15) / 15), so 0 = at a landmark. Dense distance supervision gives every point a training signal — this proved far more trainable than the sparse Gaussian heatmaps used in early experiments.
4. Post-processing — points below a per-class distance threshold are clustered with DBSCAN; each cluster contributes one landmark (the point minimizing distance-to-centroid + 2× predicted distance). Per-class (threshold, eps, min_samples) were tuned per arch and ship in config/.

Two models are provided, one per dental arch (weights/lower, weights/upper). Each has 6.45 M parameters (~26 MB). Training used MSE loss with gradient clipping, LR warmup → plateau → cosine decay (AdamW, peak 2e-4), and rigid-transform augmentation (rotation ±0.5 rad, translation ±5 mm, scale 0.8–1.2) on an NVIDIA A100.

Results (official 3DTeethLand'24 protocol)

Scored with the challenge's official evaluation code on the 50-scan test split (per-arch):

Lower-arch mean AP/AR at selected matching thresholds:

Inference takes ~0.9 s per scan on an Apple M3 Pro (CPU/MPS).

Usage

pip install -r requirements.txt
python inference.py --mesh your_scan.obj --arch lower --out landmarks.json

Or from Python:

from ptv3_dental import load_model, load_dbscan_params, detect, landmarks_to_json

model  = load_model("weights/lower/best_model.pth")
params = load_dbscan_params("config/dbscan_params_lower.json")
landmarks, _, _ = detect("your_scan.obj", model, params)   # {class: [k,3] coords}

Input: OBJ/STL/PLY mesh of a single arch, in millimeters, in the native intraoral-scanner coordinate frame (as in Teeth3DS). Output: landmark coordinates per class, in the 3DTeethLand annotation JSON format.

Limitations

• Not a medical device. Research prototype only; not validated for clinical use.
• Precision degrades below ~1 mm matching tolerance (see table) — adequate for initial planning tolerances, not yet for sub-millimeter applications.
• Supervision uses Euclidean (not geodesic) distance maps; landmarks on adjacent teeth that are close in space but far along the surface may blur together.
• Predictions vary slightly with the random point sample (controllable via seed); multi-sample consensus is a natural extension.
• Trained on 120 scans/arch from a single dataset; robustness to other scanners, partial arches, or heavy dental work is untested.

Training data & license

Trained on the 3DTeethLand'24 landmark annotations over Teeth3DS (3DTeethSeg'22) meshes — 120 training / 50 test scans per arch. The datasets are released for non-commercial research and are not redistributed here; obtain them from the challenge organizers. Weights are released under CC BY-NC 4.0 accordingly.

Links

• Code & experiments: github.com/kartikbathla/NCDS-teeth-segmentation-and-landmark-detection
• Challenge: 3DTeethLand @ MICCAI 2024
• Point Transformer V3: arXiv:2312.10035

Citation

@misc{bathla2025ptv3dental,
  author = {Bathla, Kartik},
  title  = {PTv3 Dental Landmark Detector: Direct Point-Cloud Landmark
            Detection on 3D Intraoral Scans},
  year   = {2025},
  note   = {Research internship, National Dental Centre of Singapore},
  url    = {https://huggingface.co/kartikbathla/ptv3-dental-landmark-detector}
}