Research map / 3D vision / Face and shape
3D Dental Model Segmentation with Geometrical Boundary Preserving
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 CrossTooth, which combines curvature-guided selective QEM downsampling of intraoral scan meshes with multi-view rendered image features (PSPNet) fused into a Point Transformer, to better segment tooth-gingiva boundaries.
Problem
Accurate tooth segmentation from 3D intraoral scan meshes, especially at the crown-gingiva boundary, where uniform mesh downsampling blurs geometric detail.
Approach
A curvature-weighted QEM downsampling that preserves tooth-boundary triangles, plus multi-view rendered images under vertical parallel lighting whose PSPNet features are projected back onto the points and fused with point features, with an added boundary-level contrastive (CBL) loss.
Key results
- CrossTooth reaches 95.86 mIoU and 82.05 boundary IoU on the 3DTeethSeg'22 dataset, versus 93.55 mIoU and 76.39 boundary IoU for ToothGroupNet.
- Selective downsampling increases boundary point density by roughly 10-15% over QEM (Table 1: upper jaw distance 5.029e-3 vs 5.655e-3).
- Adding image features improves CrossTooth over the point-only variant from 95.12 to 95.86 mIoU and 81.57 to 82.06 boundary IoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image features to the point network (ablation) | CrossTooth · MICCAI Challenge test set | mIoU: 95.119 -> 95.860 | Helped |
| add image features to the point network (ablation) | CrossTooth · MICCAI Challenge test set | Boundary IoU: 81.572 -> 82.058 | Helped |
| use image features only instead of fusing image and point features (ablation) | CrossTooth · MICCAI Challenge test set | mIoU: 95.860 -> 89.488 | Hurt |
| use selective downsampling instead of the alternative sampling setting (ablation) | TSGCNet · 3DTeethSeg’22 | mIoU: 89.86 -> 91.22 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | TSGCNet · 3DTeethSeg’22 | Boundary IoU: 70.33 -> 73.40 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | HiCANet · 3DTeethSeg’22 | mIoU: 90.15 -> 91.47 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | HiCANet · 3DTeethSeg’22 | Boundary IoU: 72.67 -> 75.48 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | CrossTooth · 3DTeethSeg’22 | mIoU: 93.88 -> 95.86 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | CrossTooth · 3DTeethSeg’22 | Boundary IoU: 80.73 -> 82.05 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | TSGCNet · 3D-IOSSeg | mIoU: 75.02 -> 76.45 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | TSGCNet · 3D-IOSSeg | Boundary IoU: 52.49 -> 59.92 | Helped |
| use selective downsampling instead of the alternative sampling setting (ablation) | HiCANet · 3D-IOSSeg | mIoU: 76.43 -> 78.77 | Helped |
Surprising findings
- Image-only variant reaches 89.49 mIoU, and the image contribution to the full model is modest (about 0.7 mIoU), which the authors attribute to the simple MLP fusion.
- On 3D-IOSSeg, boundary IoU declines for some methods when the number of rendered images increases beyond a certain point.
Limitations
- Performance drops on cases with few teeth or severe tooth loss, where incorrect boundary predictions occur.
- Wisdom teeth have varying appearance and few samples, leading to most wrong predictions and low accuracy.
- Feature fusion uses only a simple MLP at the last stage, which may not be fine-grained enough locally.
Experimental setup
Datasets: 3DTeethSeg'22 (MICCAI challenge, 1800 upper/lower intraoral scans) 3D-IOSSeg
Benchmarks: 3DTeethSeg'22 3D-IOSSeg
Metrics: mIoU boundary IoU average boundary point distance
Baselines: MeshSegNet SimpSegNet TSegNet TeethGNN ToothGroupNet UpToothSeg DilatedSegNet TSGCNet HiCANet
Builds on: QEM Point Transformer PSPNet Contrastive Boundary Learning ToothGroupNet
Closest papers by meaning
- Segmenting Maxillofacial Structures in CBCT Volumes (CVPR 2025 · similarity 0.85)
- NeBLa: Neural Beer-Lambert for 3D Reconstruction of Oral Structures from Panoramic Radiographs (AAAI 2024 · similarity 0.84)
- 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-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.82)
- Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.82)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 2025 · similarity 0.81)
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.