CrownGen: Patient-customized Crown Generation via Point Diffusion Model

TL;DR

CrownGen introduces a patient-customized crown generation framework that leverages a tooth-level point-cloud diffusion model conditioned on explicit cylindrical boundaries. By decoupling localization (boundary prediction) from shape synthesis (diffusion with Distance-weighted Inter-Tooth Attention), it can generate multiple, anatomically coherent crowns in a single inference pass and reproduce them as watertight meshes for CAD/CAM workflows. Quantitative benchmarks on 496 external scans and a clinical study with 26 restorations show CrownGen outperforms three state-of-the-art baselines in geometric fidelity while reducing active design time, and a formal non-inferiority analysis confirms clinical quality is not inferior to expert manual Crown design. The approach promises scalable, cost-reducing access to high-quality dental restorations and integrates smoothly with existing dental CAD platforms.

Abstract

Digital crown design remains a labor-intensive bottleneck in restorative dentistry. We present \textbf{CrownGen}, a generative framework that automates patient-customized crown design using a denoising diffusion model on a novel tooth-level point cloud representation. The system employs two core components: a boundary prediction module to establish spatial priors and a diffusion-based generative module to synthesize high-fidelity morphology for multiple teeth in a single inference pass. We validated CrownGen through a quantitative benchmark on 496 external scans and a clinical study of 26 restoration cases. Results demonstrate that CrownGen surpasses state-of-the-art models in geometric fidelity and significantly reduces active design time. Clinical assessments by trained dentists confirmed that CrownGen-assisted crowns are statistically non-inferior in quality to those produced by expert technicians using manual workflows. By automating complex prosthetic modeling, CrownGen offers a scalable solution to lower costs, shorten turnaround times, and enhance patient access to high-quality dental care.

Figures (13)

• Figure 1: Visual comparison between ground-truth natural teeth and crowns generated by CrownGen. Each row corresponds to a simulated clinical case involving multiple missing teeth for restoration. From top to bottom, scenarios shown are the restoration of a maxillary posterior sextant (FDI 12–17), mandibular anterior teeth (FDI 33–43), adjacent mandibular premolars (FDI 34–35), and adjacent mandibular molars (FDI 36–37). The rightmost column presents a surface distance error map, visualizing the point-to-surface deviation between the generated crown and the ground-truth mesh. CrownGen consistently generates anatomically plausible crowns that closely emulate the morphology of the original, healthy dentition.
• Figure 2: Geometric fidelity of generated point cloud crowns.a. Mean performance (curves) with 95% CIs (shaded bands; nonparametric bootstrap, $10,000$ resamples) of CrownGen, its ablated variants, and state-of-the-art methods across scenarios constructed from 496 test dentitions. Performance is aggregated across all scenarios, encompassing all tooth types and stratified by the number of missing teeth restored. Metrics annotated with ↓ indicate lower is better, and vice versa. Chamfer distance and Earth mover’s distance are scaled by $10^{3}$. All pairwise differences relative to CrownGen are significant (two-sided paired $t$-test; $p<0.01$). Performance stratified by tooth functional group is shown in Supplementary Figure \ref{['fig_pointcloud_toothgroup']}. Sample sizes for each condition are detailed in Supplementary Table \ref{['tab_pointcloud_all']}. b. Visual comparison of generated point clouds for five representative restoration scenarios (top to bottom): single posterior molar (FDI 16); adjacent mandibular molars (FDI 46, 47); contiguous anterior–premolar group (FDI 22–25); bilateral, inter-arch restoration (FDI 13, 43, 12, 42, 41); and four anatomically dispersed teeth (FDI 15, 11, 21, 24). While comparing methods yield plausible single-tooth shapes, they fail to produce distinct, anatomically coherent structures in multi-tooth settings. In contrast, CrownGen consistently generates discrete, well-formed crowns that respect individual tooth morphology.
• Figure 3: Ablation study on final reconstructed mesh quality. Performance comparison of CrownGen against its three ablated variants on the final reconstructed mesh surfaces of individual restored teeth across scenarios derived from 496 test dentitions. Metrics marked with ↓ indicate lower is better, and vice versa. Statistical differences relative to CrownGen were assessed using two-sided paired $t$-tests ($^{\ast\ast} p < 0.01$). Detailed sample sizes are provided in Supplementary Table \ref{['tab_ablation_metric']}.
• Figure 4: Schematic of the clinical reader study design. For each restoration case, a prepared dental scan was used to generate a digital crown via two parallel workflows: the conventional, fully manual CAD workflow and the CrownGen-assisted CAD workflow. The resulting paired designs were then presented to two trained clinical dentists for a comparative quality assessment.
• Figure 5: Overview of reader study results.a. Case-level average scores $(n=26)$ by workflow for each criterion and composite endpoints. For each case, the two readers’ scores were averaged per criterion; the composite is the per-case mean across the four criteria. Error bars show 95% CIs (case-level, nonparametric bootstrap, $10,000$ resamples). Between-workflow differences were tested with a two-sided paired $t$-test. b. Stacked probability bars of ordinal scores (1–3) by criterion $\times$ workflow; each criterion bar aggregates 52 scores (26 cases $\times$ 2 readers). The aggregate endpoint pools all criteria (208 scores per workflow). c. Design time by workflow (mean across cases). Error bars show 95% CIs. A one-sided paired $t$-test assesses faster design time ($^{\ast\ast}p<0.01$) of CrownGen-assisted CAD workflow over a fully manual, expert-led CAD workflow. d. Forest plot of paired mean differences in score by criterion and composite. Points mark the sample paired mean difference $\bar{d}$; whiskers show 95% CIs from nonparametric, case-level bootstrap CI (primary) and parametric paired $t$-based CI (sensitivity). The pre-specified non-inferiority margin is overlaid; CIs lying entirely above this margin indicate non-inferiority of the CrownGen-assisted CAD workflow compared to the fully manual, expert-led CAD workflow. e. Visual comparison of the patients' initial dentitions with the final digital crown designs generated by the CrownGen-assisted and fully manual workflows. The top two rows present external views, and the bottom row provides a cross-sectional view to illustrate internal adaptation and the occlusal contact scheme. Visualizations were created using the 3Shape 3D Viewer (version 1.3).