Intra-Oral 3D Scans: Acquisition & Analysis

• Intra-oral 3D scans are digital, high-resolution representations of a patient’s dentition and soft tissue, forming the basis for computer-aided dentistry.
• They are acquired using state-of-the-art optical scanners (structured light or confocal microscopy) that generate dense meshes with sub-100 μm accuracy.
• Automated analysis of these scans supports segmentation, instance detection, and landmarking, which enhances clinical workflows and treatment planning.

Intra-oral 3D scans are digital, high-resolution representations of the dentition and surrounding soft tissue, acquired directly from the patient's oral cavity using non-invasive optical scanners. They constitute the foundational data type for computer-aided dentistry (CAD), serving as the substrate for automatic analysis tasks central to orthodontics, prosthodontics, and diagnostics. These scans enable workflows such as segmentation, labeling, landmark detection, and multimodal integration, facilitating the automation and standardization of digital dental treatments across clinical applications (Ben-Hamadou et al., 2022).

1. Principles of Intra-Oral 3D Scan Acquisition and Annotation

State-of-the-art intra-oral scanners (e.g., Primescan, Trios3, iTero Element 2 Plus) employ structured light or confocal microscopy to generate dense 3D meshes with native surface accuracy in the 10–90 μm range and point densities of 30–80 pts/mm² (Ben-Hamadou et al., 2022). Scans are acquired for the upper and lower dental arches, with each mesh exported in standard formats (STL, OBJ, PLY) preserving both topological connectivity and per-vertex geometric attributes.

Professional annotation protocols assign each mesh vertex an FDI-based label (FDI: ISO 3950 index; 0 = gingiva, 11…48 = tooth codes) and an instance ID for tooth segmentation. Manual validation by experienced orthodontists and surgeons (>5 years’ experience) ensures clinical reliability. Mesh cleaning (removal of degenerate faces and duplicate vertices), PCA-based occlusal alignment, and optional UV-parameterization for flattened boundary annotation are used for preprocessing (Ben-Hamadou et al., 2022).

2. Core Analytical Tasks and Evaluation Metrics

Automated analysis of intra-oral 3D scans spans several hierarchical tasks, each evaluated with rigorous quantitative metrics:

• Teeth segmentation: Binary per-vertex (gingiva vs. tooth) or multi-class labeling (FDI tooth codes). Metrics: Intersection over Union (IoU), F1-score, precision/recall.
• Teeth detection (instance segmentation): Identification of individual tooth crowns as separate instances. Metric: Detection accuracy (correctly found crowns / ground truth).
• Teeth labeling: Assigning correct FDI code to each detected crown. Metric: Labeling accuracy (correctly labeled / localized instances).
• Landmark identification: Localization of clinically relevant surface points (e.g., cusp tips, fissures). Metric: Mean Euclidean distance from predicted to ground-truth landmark.

Surface accuracy is further quantified by Mean Average Symmetric Surface Distance (ASSD), enforcing sub-millimeter precision at anatomical boundaries (Ben-Hamadou et al., 2022). For landmark detection, mean error per landmark and percent within threshold (e.g., 1 mm, 2 mm) are used (Ben-Hamadou et al., 9 Dec 2025, Wu et al., 2021).

3. Segmentation and Labeling: Architectures and Baselines

Methods span multiple deep learning paradigms, optimized for the unique challenges of dental mesh geometry:

• Point-cloud networks: PointNet/PointNet++ and MeshSNet learn per-point features on unordered point sets and mesh surfaces, respectively (Ben-Hamadou et al., 2022).
• Graph-based CNNs: Use k-NN or mesh adjacency to propagate context and attention across the dental arch (e.g., Graph Attention Convolution (Ben-Hamadou et al., 2022), DGCNN (Hao et al., 2022)).
• Voxel-based 3D CNNs: Adaptations of U-Net handle voxelized inputs but are limited by dental surface resolution and memory load (Ben-Hamadou et al., 2022).

Transformers and hybrid architectures (EdgeConv + Transformer), such as TSegFormer and TFormer, have demonstrated state-of-the-art segmentation performance by capturing both local geometric cues (via curvature and EdgeConv) and global tooth–tooth interactions using multi-head self-attention (Xiong et al., 2023, Xiong et al., 2022). Geometry-guided loss functions focus learning at sharp boundaries, directly addressing common segmentation failure modes at tooth–gingiva and inter-tooth crests (Xiong et al., 2023). Multi-task heads (tooth vs. gingiva) further disambiguate boundary regions.

Recent methods systematically outperform earlier baselines: TSegFormer achieves mean IoU of 94.3% and Dice of 96.0% on 16,000-scan datasets, surpassing edge baselines (PointNet++, mIoU 82.6%) (Xiong et al., 2023). Instance-level crowns can be detected with up to 96.6% accuracy on complete arches, but performance degrades in partial scans or cases of severe crowding (Ben-Hamadou et al., 2022, Jana et al., 2023).

4. Challenges in Partial Scans and Boundary Delineation

Most methods assume full-jaw scans. Performance collapses when applied to partial scans (e.g., single quadrants or isolated teeth), with Dice dropping by 40–65% (PointMLP, MeshSegNet+GCO) (Jana et al., 2023). Key causes include loss of arch context, boundary artifacts at crop edges, and resolution mismatch in graph construction.

Approaches to mitigate these issues include:

• Partial-crop data augmentation during training, upsampling partial meshes to models’ input sizes.
• Multi-resolution and sliding window inference to preserve local features while accommodating variable scan extents.
• Integration of anatomical priors and transformer-based attention to compensate for missing context.
• Boundary-aware regularization to penalize label discontinuities only at anatomical—not artificial—margins.

Boundary precision, especially at the gingiva-enamel junction, is critical for prosthetic and periodontal planning. CrossTooth, for example, applies curvature-weighted mesh downsampling and multi-view image fusion (via PSPNet) to enhance boundary recall (+5.7 pp boundary IoU over previous best) (Xi et al., 31 Mar 2025).

5. Landmark Identification and Clinical Applications

Automatic detection of anatomical landmarks (e.g., cusps, contact points, fossae) underpins diagnostic indices and orthodontic device placement. The 3DTeethLand challenge established the first annotated dataset for this task, with structures labeled by consensus of multiple clinicians (Ben-Hamadou et al., 9 Dec 2025). Leading pipeline architectures employ initial tooth segmentation, followed by local landmark detection (e.g., PointNet-Reg heatmap regression per tooth ROI), achieving sub-millimeter mean localization errors (≈0.6 mm) (Wu et al., 2021).

Multi-stage frameworks combining mesh segmentation, instance cropping, and spatially aware point-based networks (e.g., TeethGNN, PointTransformer V3, TL-DETR) allow the modular detection and measurement of clinically meaningful coordinates at scale (Ben-Hamadou et al., 9 Dec 2025). Success rates at 1 mm error exceed 85% for most landmark types on current benchmarks.

Such pipeline-generated landmarks can be directly ingested by CAD/CAM or orthodontic simulation systems for virtual bracket planning, quantitative movement assessment, or arch shape synthesis.

6. Integration with Multimodal and Clinical Workflows

High-fidelity intra-oral scans are increasingly combined with CBCT (for root and bone imaging), cephalometric radiographs (for 2D/3D alignment), and facial scans for advanced diagnostics and appliance design. Standardized alignment frameworks (e.g., U-Midline Dental Axis, surface-based DRR) allow robust, pose-invariant registration of intraoral meshes with other modalities (Miao et al., 18 Nov 2025). In full-jaw reconstructions, implicit SDF fusion models yield watertight, anatomically correct whole-tooth models suitable for CAD/CAM and surgical planning, with errors <0.1 mm in the anatomical crown (Zhu et al., 21 Jan 2026).

Automated intra-oral scan analysis enables:

• Virtual orthodontic setups, aligner and fixed appliance planning.
• Custom-fitted prosthetics with precise gingival or margin delineations.
• Personalization of occlusal splints and jaw-repositioning devices via direct mesh manipulation and registration (Tomaka et al., 17 Apr 2025).
• Population-level analysis for oral health epidemiology.

7. Future Directions and Open Problems

Major research challenges remain:

• Robust instance segmentation in the presence of severe crowding, missing teeth, or appliances (implants, brackets, prosthetics).
• Boundary refinement at the gingiva–enamel interface, especially in cases of disease, malocclusion, or complex tissue morphology.
• Generalization to partial-scans and variable field-of-view acquisition.
• Domain adaptation to unseen scanners, populations, or pathologies.
• Extension to richer annotation targets, such as precise morphometric landmarks or pathologic features.

Ongoing work seeks semi-supervised and self-supervised paradigms to reduce annotation cost, transformer-based multi-task models for joint segmentation and landmarking, and real-time inference optimized for clinical chairside deployment (Xiong et al., 2023, Xiong et al., 2022, Ben-Hamadou et al., 9 Dec 2025). The development and broad public release of large expert-validated datasets such as Teeth3DS+ and 3DTeethLand is expected to remain foundational in driving progress and translating algorithmic advances to clinical dental CAD systems (Ben-Hamadou et al., 2022, Ben-Hamadou et al., 9 Dec 2025).

---

References:

• Teeth3DS+: An Extended Benchmark for Intraoral 3D Scans Analysis (Ben-Hamadou et al., 2022)
• 3DTeethSeg’22: 3D Teeth Scan Segmentation and Labeling Challenge (Ben-Hamadou et al., 2023)
• Two-Stage Mesh Deep Learning for Automated Tooth Segmentation and Landmark Localization on 3D Intraoral Scans (Wu et al., 2021)
• 3D Tooth Segmentation in Intraoral Scans with Geometry Guided Transformer (Xiong et al., 2023)
• A Critical Analysis of the Limitation of Deep Learning based 3D Dental Mesh Segmentation Methods in Segmenting Partial Scans (Jana et al., 2023)
• 3DTeethLand challenge (Dental Landmark Detection) (Ben-Hamadou et al., 9 Dec 2025)
• CrossTooth boundary-preserving segmentation (Xi et al., 31 Mar 2025)
• High-Fidelity 3D Tooth Reconstruction by Fusing Intraoral Scans and CBCT Data via a Deep Implicit Representation (Zhu et al., 21 Jan 2026)
• Silhouette-to-Contour Registration for 3D-2D Alignment (Miao et al., 18 Nov 2025)