Research map / Video / Video generation
Collaborative Tooth Motion Diffusion Model in Digital Orthodontics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a conditional diffusion model with a tooth latent representation (tooth frame plus shape code), a GRU denoiser, graph-based multi-tooth interaction, and collaborative losses to generate intermediate tooth motion sequences in orthodontics.
Problem
Generating the full intermediate tooth motion process from initial malocclusion to target ideal alignment, while avoiding collisions and respecting occlusal relationships between upper and lower teeth.
Approach
Remodels tooth motion planning as conditional diffusion-based distribution fitting: each tooth gets a diffusion denoiser conditioned on a compact tooth latent (frame plus shape code), with a gated graph network for inter-tooth interaction and frame, point-cloud (occlusal), and Lennard-Jones collision losses as soft constraints.
Key results
- Proposed method achieves Rerror 0.18 deg, Terror 0.039 mm and AUCpoint 93.06 at 20 steps, versus 0.29/0.074/84.95 for interpolation and 0.44/0.043/86.53 for regression.
- MDM baseline at 20 steps has AUCpoint 75.39 and Rerror 0.46, and the paper reports serious collisions for it.
- Post-processing with SDF-based collision control further lowers Terror to 0.0386 mm in the ablation table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use collaborative tooth motion diffusion instead of interpolation | Tooth motion generation · Tooth motion test set, 20 steps | AUCpoint: 84.95 -> 93.06 | Helped |
| Use collaborative tooth motion diffusion instead of TANet | Tooth motion generation · Tooth motion test set, 5 steps | Rerror: 3.24 -> 0.36 | Helped |
| Use collaborative tooth motion diffusion instead of regression | Tooth motion generation · Tooth motion test set, 20 steps | AUCpoint: 86.53 -> 93.06 | Helped |
| Use collaborative tooth motion diffusion instead of MDM | Tooth motion generation · Tooth motion test set, 20 steps | AUCpoint: 75.39 -> 93.06 | Helped |
| Add tooth frame conditioning (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 65.1907 -> 90.9010 | Helped |
| Add tooth shape code (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 90.9010 -> 92.2750 | Helped |
| Add multi-tooth interaction module (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 90.9010 -> 92.1020 | Helped |
| Add frame constraint loss (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 90.9010 -> 92.7169 | Helped |
| Add tooth and jaw constraint loss (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 90.9010 -> 92.6158 | Helped |
| Add collision-avoiding loss (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 90.9010 -> 92.0803 | Helped |
| Add all proposed components and constraints (ablation) | Base diffusion network · Tooth motion test set, 20 steps | AUCpoint: 90.9010 -> 93.0670 | Helped |
| Apply post-processing collision correction (ablation) | Our · Tooth motion test set, 20 steps | AUCpoint: 93.0670 -> 93.2190 | Helped |
Surprising findings
- Plain MDM-style image-inpainting diffusion performed worse than simple regression at 20 steps on AUCpoint (75.39 vs 86.53).
Limitations
- Collision-avoidance is a soft constraint in training, so outputs may still have slight collisions and require post-processing.
- Evaluated only on a private dataset of 1050 dental cases, not publicly released.
Experimental setup
Datasets: in-house dental case dataset from hospitals (1050 cases, dentist-labeled tooth motion processes)
Benchmarks: in-house tooth motion test set (158 cases)
Metrics: Rerror Terror AUCpoint
Baselines: Interpolation TANet Regression MDM
Builds on: Diffusion probabilistic models DDIM MDM TANet TAD-Net
Closest papers by meaning
- Teeth-SEG: An Efficient Instance Segmentation Framework for Orthodontic Treatment based on Multi-Scale Aggregation and Anthropic Prior Knowledge (CVPR 2024 · similarity 0.82)
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 2026 · similarity 0.81)
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.80)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.80)
- Constrained Synthesis with Projected Diffusion Models (NeurIPS 2024 · similarity 0.80)
- Single Motion Diffusion (ICLR 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.