Research map / Detection and segmentation / Open-set and few-shot detection
Segmenting Maxillofacial Structures in CBCT Volumes
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.
Introduces ToothFairy2, a public CBCT dataset with 3D voxel-level annotations of 42 maxillofacial classes, and benchmarks CNN, transformer, and Mamba segmentation models plus nnU-Net adaptations.
Problem
Maxillofacial structure segmentation in CBCT volumes is labor-intensive, and existing datasets are private, small, or cover only a few structures and lack 3D annotations.
Approach
A large publicly accessible CBCT dataset with 42 3D-annotated classes, produced via iterative model-assisted annotation with clinician refinement and alpha-shape smoothing, plus nnU-Net modifications (no left/right mirroring, deeper bottleneck, connected-component post-processing) that improve benchmark performance.
Key results
- Swin-UMamba, UMamba, and other baselines were benchmarked; UMamba achieved the best default-plan average DSC of 85.05 among baselines.
- Removing left/right mirroring and increasing depth, plus post-processing, raised nnU-Net ResEnc from 74.16 to 84.99 DSC on the test set.
- With the adapted plan (w/o post-processing), UNETR++ improved from 71.43 to 83.57 DSC and UMamba reached 85.39 DSC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UMamba instead of the other evaluated architectures | ToothFairy2 segmentation · ToothFairy2 test | DSC: 85.05 | Helped |
| Disable left/right mirroring (ablation) | nnU-Net ResEnc · ToothFairy2 test | DSC: 74.16 -> 80.79 | Helped |
| Disable left/right mirroring (ablation) | nnU-Net ResEnc · ToothFairy2 test | HD95: 14.48 -> 12.37 | Helped |
| Increase network depth after disabling left/right mirroring (ablation) | nnU-Net ResEnc · ToothFairy2 test | DSC: 80.79 -> 82.11 | Helped |
| Increase network depth after disabling left/right mirroring (ablation) | nnU-Net ResEnc · ToothFairy2 test | HD95: 12.37 -> 11.86 | Helped |
| Add connected-component post-processing (ablation) | nnU-Net ResEnc without left/right mirroring and with increased depth · ToothFairy2 test | DSC: 82.11 -> 84.99 | Helped |
| Add connected-component post-processing (ablation) | nnU-Net ResEnc without left/right mirroring and with increased depth · ToothFairy2 test | HD95: 11.86 -> 8.57 | Helped |
| Use the personalized nnU-Net plan (ablation) | nnU-Net ResEnc · ToothFairy2 test | DSC: 74.16 -> 82.11 | Helped |
| Use the personalized nnU-Net plan (ablation) | UMamba · ToothFairy2 test | DSC: 85.05 -> 85.39 | No clear effect |
| Use the personalized nnU-Net plan (ablation) | UMamba · ToothFairy2 test | HD95: 5.28 -> 4.70 | Helped |
| Train on ToothFairy2 rather than Cui data | UMamba · Our test set, teeth classes | DSC: 79.95 -> 90.18 | Helped |
Surprising findings
- 2D models (VMamba, Swin-UMamba) outperformed intrinsically 3D models on many structure groups.
- Left/right mirroring augmentation hurt performance because maxillofacial structures are bilaterally symmetric and left/right distinction matters.
Limitations
- Training data come from a single acquisition center, limiting generalizability.
- Implants, crowns, and bridges ('others' classes) are underrepresented, lowering performance on them.
- Some maxillofacial components such as braces and bone plates are absent from the cohort.
Experimental setup
Datasets: ToothFairy2 (480 training CBCT scans) Cui et al. CBCT tooth dataset
Benchmarks: ToothFairy2 test set (50 CBCT scans, grand-challenge) Cui et al. CBCT dataset
Metrics: DSC HD95
Baselines: nnU-Net nnU-Net ResEnc TransUNet nnFormer UNETR++ UMamba VMamba Swin-UMamba
Builds on: nnU-Net U-Mamba UNETR++ nnFormer VMamba
Closest papers by meaning
- Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.85)
- 3D Dental Model Segmentation with Geometrical Boundary Preserving (CVPR 2025 · similarity 0.85)
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.85)
- Teeth-SEG: An Efficient Instance Segmentation Framework for Orthodontic Treatment based on Multi-Scale Aggregation and Anthropic Prior Knowledge (CVPR 2024 · similarity 0.85)
- NeBLa: Neural Beer-Lambert for 3D Reconstruction of Oral Structures from Panoramic Radiographs (AAAI 2024 · similarity 0.85)
- C^2RV: Cross-Regional and Cross-View Learning for Sparse-View CBCT Reconstruction (CVPR 2024 · similarity 0.84)
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.