Research map / Detection and segmentation / CNN features
Teeth-SEG: An Efficient Instance Segmentation Framework for Orthodontic Treatment based on Multi-Scale Aggregation and Anthropic Prior Knowledge
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 TeethSEG, a ViT-based (CLIP-L/14) instance segmentation framework for 2D intraoral teeth with multi-scale aggregation blocks, permutation-based upscalers, cross-gating, and an anthropic prior knowledge layer, plus a new 150K-image dataset IO150K.
Problem
Teeth localization, segmentation and labeling in 2D intraoral images is hard for general instance/semantic segmentation models due to similar tooth shapes, position variation across subjects, and abnormalities such as caries and edentulism.
Approach
A transformer with trainable class tokens for foreground, background and 16 tooth IDs; cross/self-gating (element-wise importance scaling instead of weighted-sum attention) keeps tooth-token embeddings distinct; a permutation-based upscaler replaces bilinear interpolation to get sharp edges; and an APK layer masks self-gating so that only adjacent and contralateral teeth interact, encoding orthodo
Key results
- TeethSEG reaches 0.91 mIoU on the i.i.d. test, about 4 points above the best competitor (ViT-Adapter-L, 0.87).
- On the RGB test split TeethSEG gets 0.91 mIoU versus 0.85 for ViT-Adapter-L after fine-tuning.
- Ablation: removing APK drops o.o.d. mIoU from 0.84 to 0.76 and removing MSA drops it to 0.73 (variant c/d comparisons).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TeethSEG instead of ViT-Adapter-L | IO150K segmentation · IO150K i.i.d. test | mIoU: 0.87 -> 0.91 | Helped |
| use TeethSEG instead of Segformer | IO150K segmentation · IO150K o.o.d. test | mIoU: 0.81 -> 0.84 | Helped |
| use TeethSEG instead of ViT-Adapter-L | IO150K RGB segmentation · IO150K RGB test | mIoU: 0.85 -> 0.91 | Helped |
| use permutation-based upscalers instead of bilinear interpolation (ablation) | TeethSEG · IO150K i.i.d. test | mIoU: 0.72 -> 0.91 | Helped |
| use permutation-based upscalers instead of bilinear interpolation (ablation) | TeethSEG · IO150K o.o.d. test | mIoU: 0.67 -> 0.84 | Helped |
| use cross/self-gating instead of cross/self-attention (ablation) | TeethSEG · IO150K i.i.d. test | mIoU: 0.89 -> 0.91 | Helped |
| use cross/self-gating instead of cross/self-attention (ablation) | TeethSEG · IO150K o.o.d. test | mIoU: 0.80 -> 0.84 | Helped |
| add MSA blocks (ablation) | TeethSEG · IO150K i.i.d. test | mIoU: 0.87 -> 0.91 | Helped |
| add MSA blocks (ablation) | TeethSEG · IO150K o.o.d. test | mIoU: 0.73 -> 0.84 | Helped |
| add APK layer (ablation) | TeethSEG · IO150K i.i.d. test | mIoU: 0.89 -> 0.91 | Helped |
| add APK layer (ablation) | TeethSEG · IO150K o.o.d. test | mIoU: 0.76 -> 0.84 | Helped |
| add MSA blocks and APK layer (ablation) | TeethSEG · IO150K i.i.d. test | mIoU: 0.79 -> 0.91 | Helped |
Surprising findings
- Bilinear interpolation upsampling in transformer decoders causes mesh-like edge artifacts, which the permutation upscaler avoids.
Limitations
- Code and dataset are promised ('will be available') but no release is confirmed in the text.
- Evaluation is on the authors' own dataset; the o.o.d. and RGB splits are relatively small (RGB0.8K).
Experimental setup
Datasets: IO150K Challenge80K Plaster70K RGB0.8K
Benchmarks: IO150K i.i.d. test IO150K o.o.d. test IO150K RGB test
Metrics: mIoU per-class IoU
Baselines: DeepLabV3 Segformer Segmenter Swin-L SwinV2-G BeiT-B ViT-Adapter-L
Builds on: Segmenter CLIP ViT-Adapter SAM
Closest papers by meaning
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.86)
- Segmenting Maxillofacial Structures in CBCT Volumes (CVPR 2025 · similarity 0.85)
- OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining (CVPR 2025 · similarity 0.83)
- Seg-ReSearch: Segmentation with Interleaved Reasoning and External Search (ICML 2026 · similarity 0.83)
- 3D Dental Model Segmentation with Geometrical Boundary Preserving (CVPR 2025 · similarity 0.82)
- Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.82)
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.