CrownGen: Automated Dental Crown Design

• CrownGen is a deep generative framework that automates customized dental crown design by integrating 3D scans, spatial priors, and inter-tooth context.
• It employs architectures from conditional GANs to modern diffusion models to achieve high geometric fidelity and clinically acceptable restorations.
• Innovative modules such as geometry-aware transformers and boundary prediction enhance occlusal accuracy and scalable efficiency in restorative workflows.

CrownGen is a class of deep generative frameworks for automating the design of patient-customized dental crowns, spanning architectures from early conditional GANs to modern point diffusion models. It replaces the traditionally manual, labor-intensive process of crown sculpting with neural inference, leveraging 3D scan data, spatial priors, and explicit inter-tooth context. CrownGen achieves high geometric fidelity, non-inferior clinical quality to expert-crafted crowns, and scalable efficiency in restorative dental workflows (Hwang et al., 2018, Hosseinimanesh et al., 2023, Bae et al., 26 Dec 2025).

1. Technical Foundations and Evolution

CrownGen addresses central challenges in crown fabrication: anatomical accuracy, functional occlusion, prevention of inter-dental collisions, and realistic contact-point placement. Early systems operated on 2D depth-maps and utilized conditional generative adversarial networks (cGAN), synthesizing crown topography conditioned on missing-tooth and opposing-jaw scans augmented by inter-jaw gap statistics (Hwang et al., 2018). Modern CrownGen systems employ 3D point cloud representations and generative diffusion models, enabling multi-crown synthesis in a single pass and improved modeling of inter-tooth dependencies (Bae et al., 26 Dec 2025).

The domain has seen the integration of boundary-aware modules, geometry-aware transformer encoders for point-cloud completion, and advanced self-attention for inter- and intra-tooth context modeling (Hosseinimanesh et al., 2023, Bae et al., 26 Dec 2025). Each advance has been motivated by observed limitations in previous CAD workflows: inability to account for subtle occlusal grooves, inefficient enforcement of non-penetration constraints, and suboptimal contact-point distributions.

2. Input Representation and Data Preprocessing

Depth-Map and Point Cloud Modalities

• Depth-map input (Early CrownGen):
  • $x \in \mathbb{R}^{H \times W \times 1}$: Prepared jaw scan (missing crown region).
  • $\tilde{x} \in \mathbb{R}^{H \times W \times 1}$: Opposing jaw scan.
  • $d \in \mathbb{R}^{H \times W \times 1}$: Pixelwise gap distances (Hwang et al., 2018).
• 3D Point cloud input (Modern CrownGen):
  • Each tooth (context or target) sampled to $N$ points with features: $(x_i, y_i, z_i, b_i, t_i, B_i)$, where $b$ is a context/target indicator, $t$ is an FDI tooth index embedding, and $B$ contains cylindrical boundary priors (Bae et al., 26 Dec 2025).
  • Margin-line guidance: For transformer-based pipelines, a high-resolution spline representing the margin line (critical cementation boundary) is sampled and concatenated with the crown shell to enforce fidelity at prosthetic margins (Hosseinimanesh et al., 2023).

Preprocessing Pipeline

All approaches require robust tooth segmentation (e.g., MeshSegNet, TSegFormer), occlusal plane alignment, normalization of point clouds, and casewise extraction of neighboring and antagonistic teeth for context. Data augmentation includes random mirroring, rotation, scaling, and translation jitter, maintaining anatomical realism.

3. Architectural Overview

Approach	Core Architecture	Key Modules / Innovations
Early CrownGen (GAN)	U-Net cGAN (57M params)	Histogram loss; gap-map regularization
Transformer-based	Geometry-aware transformer	Margin-line embedding; FoldingNet decoder
CrownGen (Diffusion)	DDPM + PointNet++ PVC/DITA	Tooth-level object split; Inter-tooth attention; boundary prediction module

Boundary Prediction Module (Diffusion model version): Predicts cylindrical spatial priors for each missing tooth using a backbone of Set Abstraction (PointNet++) with Point-Voxel Convolution and Distance-weighted Inter-Tooth Attention (DITA).

Generative Module:

• cGAN: Generator receives stacked depth-maps and gap images; decoder outputs crown depth-map.
• Transformer: Encodes context+margin line, seeds proxy queries, decodes surface via multi-stage folding.
• Diffusion: Conditional DDPM on concatenated crowns, context teeth held fixed, denoises from random samples inside predicted boundaries (Bae et al., 26 Dec 2025).

DITA explicitly encodes anatomical relationships (proximal/antagonistic) between tooth objects, using FDI index differences and learned relative positional encodings.

4. Training Methodologies and Loss Functions

• Adversarial and Reconstruction Objectives:
  • cGAN adversarial loss plus L1 loss for morphological accuracy.
• Histogram/functionality Losses:
  • Regularizes the gap map to match the target's inter-jaw distances, integrating occlusal constraints and contact sparsity (with options for bin weighting and higher-order pooling).
• Chamfer and Margin-Line Losses:
  • For point clouds, Chamfer distance serves as a surrogate for geometric similarity; margin-line fidelity enforced by high-resolution sampling of the target interface.
• Diffusion Loss:
  • Mean squared error between true/predicted noise vectors in DDPM, conditioning on context, boundaries, and time step (Bae et al., 26 Dec 2025).

Training involves alternating generator/discriminator updates for GANs, Adam optimizers with tailored learning-rate schedules, dropout, and large-scale augmentation. In the latest diffusion approach, pseudo-crown self-training is used to utilize partially edentulous scans, addressing a data bottleneck for large-scale deployment.

5. Quantitative Benchmarks and Clinical Evaluation

Morphology and Functional Fit

• Depth-map GAN (validation): RMSE ≤ 0.07 mm, IOU > 91%, boundary $F_1 > 93\%$, gap penetration failures reduced from 85.6% (no context) to 7.8% (histogram-regularized).
• Transformer-based: Chamfer error reduction from $5.639\times10^{-2}$ (no margin line) to $8.04\times10^{-3}$; margin-line error (max/mean) reduced by ~4–6× with margin-line input (Hosseinimanesh et al., 2023).
• Diffusion model (point clouds):
  • Chamfer Distance (CD): $34.53\times10^{-3}$ vs. $38.78\times10^{-3}$ for prior SOTA (PointSea), stable with increasing $k$ (number of missing teeth), while baselines degrade.
  • F1-score @ 0.3 mm: 0.471 (CrownGen) vs. 0.466 (PointSea) for $k=1$.
  • Multi-tooth completion shows robust CD/EMD across $k=1$ to 6, unlike baselines (Bae et al., 26 Dec 2025).

Clinical Study

• Active Design Time: CrownGen-assisted design requires $740\pm131$ s, reducing time by $17.8\%$ relative to manual workflows (900±180 s, $p<0.01$).
• Quality Assessment: Composite clinical score $2.938\pm0.194$ (CrownGen) vs. $2.928\pm0.197$ (manual), non-inferior by pre-specified margin, with no significant differences across occlusion, contact, alignment, contour.
• Inter-rater Reliability: Gwet’s AC$_2$ = 0.947, percent agreement 89.4%.

6. Key Contributions and Comparative Innovations

• Tooth-object Decomposition: Enables variable-cardinality multi-crown completion without combinatorial complexity, facilitating efficient inference across edentulous scenarios (Bae et al., 26 Dec 2025).
• Inter-tooth Context Modeling: DITA modules substantially improve geometric and functional plausibility (ablation: removing DITA or boundaries degrades CD by 17–29%, F1 by 30–57%).
• Margin-line Emphasis: High-resolution boundary point upweighting in the transformer approach ensures fit, sealing, and clinical cementation precision (Hosseinimanesh et al., 2023).
• Pseudo-crown Self-bootstrapping: Allows semi-supervised expansion of training on partially edentulous scans—critical for real-world generalization (Bae et al., 26 Dec 2025).

7. Limitations and Future Directions

• Dimensionality: Early approaches are depth-map based; full 3D volumetric/multi-view extensions are needed for highly complex morphologies (Hwang et al., 2018).
• Runtime: Diffusion inference ($\sim$85 s on RTX 4090) is longer than feed-forward models but minor compared to manual CAD steps (Bae et al., 26 Dec 2025).
• Margin Adaptation: Latest CrownGen models generate supragingival morphology but leave final margin adaptation to post-processing CAD tools owing to the variability of clinical margin geometry (Bae et al., 26 Dec 2025).
• Potential Improvements: Faster denoising (e.g., DDIM), joint margin modeling, and uncertainty-aware context integration for damaged teeth are promising avenues.

A plausible implication is that future CrownGen iterations may entirely supplant the manual morphology-driven phase of crown design, reserving technician input for final margin and esthetic adjustments.

---

CrownGen marks a paradigm shift in digital prosthodontics, providing a rigorously-validated, data-driven framework that automates key phases of patient-customized crown design with accuracy, efficiency, and scalability unmatched by traditional manual or CAD-assisted workflows (Hwang et al., 2018, Hosseinimanesh et al., 2023, Bae et al., 26 Dec 2025).