MADCrowner: Margin Aware Dental Crown Design with Template Deformation and Refinement

Linda Wei; Chang Liu; Wenran Zhang; Yuxuan Hu; Ruiyang Li; Feng Qi; Changyao Tian; Ke Wang

cs.CVMar 5, 2026

MADCrowner: Margin Aware Dental Crown Design with Template Deformation and Refinement

Authors: Linda Wei, Chang Liu, Wenran Zhang, Yuxuan Hu, Ruiyang Li, Feng Qi, Changyao Tian, Ke Wang,

Abstract

Dental crown restoration is one of the most common treatment modalities for tooth defect, where personalized dental crown design is critical. While computer-aided design (CAD) systems have notably enhanced the efficiency of dental crown design, extensive manual adjustments are still required in the clinic workflow. Recent studies have explored the application of learning-based methods for the automated generation of restorative dental crowns. Nevertheless, these approaches were challenged by inadequate spatial resolution, noisy outputs, and overextension of surface reconstruction. To address these limitations, we propose \totalframework, a margin-aware mesh generation framework comprising CrownDeformR and CrownSegger. Inspired by the clinic manual workflow of dental crown design, we designed CrownDeformR to deform an initial template to the target crown based on anatomical context, which is extracted by a multi-scale intraoral scan encoder. Additionally, we introduced \marginseg, a novel margin segmentation network, to extract the cervical margin of the target tooth. The performance of CrownDeformR improved with the cervical margin as an extra constraint. And it was also utilized as the boundary condition for the tailored postprocessing method, which removed the overextended area of the reconstructed surface. We constructed a large-scale intraoral scan dataset and performed extensive experiments. The proposed method significantly outperformed existing approaches in both geometric accuracy and clinical feasibility.

Figures & tables

Figure 1: Dental Crown Designing Procedure. With the patient’s intraoral scan, the technician utilizes a CAD system to extract the cervical margin, select and modify a digital template of the target tooth, culminating in the design of the final crown.

Figure 2: The visualization of surface reconstruction for a certain dental crown. (a) The original crown mesh. (b) point cloud with normal vectors of the crown mesh. (c) Reconstructed surface of (b) by Poisson surface reconstruction. The bottom of the reconstructed surface is compulsorily sealed due to the intrinsic constraints of the reconstruction algorithm.

Figure 3: The workflow of MADCrowner. The input IOS and the cervical margin identified by CrownSegger are sent to CrownDeformR to generate a watertight crown by deforming and refining an initial template. The extraneous regions are excised, guided by the margin line, to obtain the result.

Figure 4: Architecture of CrownSegger. The CrownSegger fuses both the point-wise and voxel-wise features and predicts the segmentation of the prepared abutment. The margin line is obtained by extracting the boundary of the segmentation mask.

Figure 5: Overall architecture of CrownDeformR. CrownDeformR consists of three components: IOS feature extraction, Crown Template Deformation, and Coarse Crown Refinement. For each generated crown mesh, we utilize the margin line identified by CrownSegger to excise the extraneous regions and solve the intrinsic watertightness issue introduced by DPSR.

Jaw Side	Tooth Position	Number of Cases
Upper	1st Premolar	543
	2nd Premolar	412
	1st Molar	1096
	2nd Molar	270
Lower	1st Premolar	152
	2nd Premolar	178

Table 1: Data Distribution for Dental Crown Design

Methods	Accuracy $\uparrow$	IoU $\uparrow$	Hausdorff Distance (mm) $\downarrow$
PointNet [ 30 ]	0.935	0.858	2.450
PointNet++ [ 31 ]	0.956	0.901	1.720
PointTransformer [ 49 ]	0.963	0.939	1.238
CrownSegger	0.991	0.972	0.328
CrownSegger(zero-shot on crown generation dataset)	–	–	0.545

Table 2: Comparison experiment on cervical margin segmentation. The best results are highlighted in blue . The CrownSegger achieves the best performance on the test set. It also exhibits excellent performance under a zero-shot setting on crown-generated preparations, achieving a Hausdorff Distance of 0.545 mm without additional finetuning.

Methods	CD-L2 ( $\text{mm}^{2}$ ) $\downarrow$			Fidelity Distance ( $\text{mm}^{2}$ ) $\downarrow$			Hausdorff Distance (mm) $\downarrow$			F-score $\uparrow$
Methods	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall
PCN [ 47 ] +SAP	0.271	0.318	0.305	0.118	0.148	0.140	1.400	1.458	1.441	0.850	0.889	0.878
TopNet [ 35 ] +SAP	0.354	0.409	0.394	0.164	0.200	0.190	1.752	1.869	1.836	0.770	0.812	0.800
GRnet [ 41 ] +SAP	0.230	0.254	0.247	0.110	0.132	0.126	1.118	1.139	1.134	0.874	0.901	0.894
DMCv2 [ 12 ]	0.227	0.264	0.254	0.107	0.132	0.125	1.227	1.252	1.245	0.868	0.897	0.888
MADCrowner	0.176	0.174	0.175	0.098	0.092	0.094	1.093	0.996	1.027	0.903	0.932	0.924

Table 3: Performance comparison of different crown generation methods before mesh reconstruction. The best and second-best results are highlighted in blue and green , respectively.

Methods	CD-L2 ( $\text{mm}^{2}$ ) $\downarrow$			Fidelity Distance ( $\text{mm}^{2}$ ) $\downarrow$			Hausdorff Distance (mm) $\downarrow$			F-score $\uparrow$
Methods	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall
GRnet+SAP [ 41 ]	0.212	0.210	0.210	0.111	0.100	0.103	1.212	1.307	1.280	0.895	0.901	0.899
DMCv2 [ 12 ]	0.259	0.251	0.253	0.137	0.123	0.127	1.329	1.319	1.322	0.868	0.875	0.873
VBCD [ 40 ]	0.212	0.207	0.209	0.108	0.109	0.109	1.115	1.163	1.150	0.896	0.915	0.909
Diffusion SDF [ 4 ]	0.228	0.215	0.219	0.116	0.108	0.110	1.216	1.458	1.390	0.875	0.900	0.893
MADCrowner	0.198	0.180	0.185	0.083	0.087	0.086	1.096	1.026	1.046	0.907	0.920	0.917

Table 4: Quantitative evaluation of the crown meshes generated by different methods. For a fair comparison, all the outputs of these methods were processed using our post-processing pipeline to remove superfluous areas resulting from surface reconstruction. The best results are highlighted in blue . To demonstrate the effect of the post-processing algorithm, the last row shows the performance of our method without the post-processing algorithm.

Figure 6: Visualization of the comparison experiment for abutment segmentation. The cervical margin identified by CrownSegger is much closer to the ground truth.

Figure 7: Comparison Result of the MADCrowner with other point cloud based methods.

Figure 8: Comparison Result of the MADCrowner with voxel based methods.

Figure 9: Visualization of the ablation study. (a) to (e) are corresponding to the results in Table 5 , which are models using: (a) a hemisphere template without the template deformation module; (b) a hemisphere template with the deformation module; (c) the initial crown template without the template deformation module; (d) no cervical margin constraint; and (e) CrownDeformR, our full model. (f) represents the ground truth.

Figure 10: point cloud visualization of the ablation study for cervical margin constraint. The red point cloud denotes the ground truth, and the blue one denotes the generation result. (a) MADCrowner without Cervical Margin Constraint. (b) MADCrowner.

Components			CD-L2 ( $\text{mm}^{2}$ ) $\downarrow$			Fidelity Distance ( $\text{mm}^{2}$ ) $\downarrow$			Hausdorff Distance (mm) $\downarrow$			F-Score $\uparrow$			Efficiency
Cervical Margin Constraints	Initial Crown Template	Crown Template Deformation	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall	VRAM (MB) $\downarrow$	Inference Time (ms) $\downarrow$
✗	✗	✗	0.214	0.222	0.220	0.106	0.114	0.112	1.419	1.347	1.367	0.866	0.905	0.895	1229	87
✗	✓	✗	0.204	0.217	0.213	0.102	0.110	0.107	1.372	1.291	1.314	0.878	0.910	0.901	1229	90
✗	✗	✓	0.215	0.211	0.212	0.115	0.106	0.108	1.365	1.280	1.301	0.882	0.911	0.903	1305	100
✗	✓	✓	0.188	0.195	0.193	0.099	0.102	0.101	1.182	1.152	1.160	0.886	0.921	0.911	1305	103
✓	✓	✓	0.176	0.174	0.175	0.098	0.092	0.094	1.093	0.996	1.027	0.903	0.932	0.924	1305	105

Table 5: Ablation study results of different components in CrownDeformR. The best results are highlighted in blue .

Figure 11: Ablation Study on Curvature Weight.

Figure 12: Proximal Contact Visualization. The color map represents the nearest distance to the adjacent teeth. (a) The generation result of a premolar of VBCD. (b) The generation result of a premolar of MADCrowner. (c) Ground truth of the premolar case. (d) The generation result of a molar of VBCD. (e) The generation result of a molar of MADCrowner. (f) Ground truth of the molar case.

Methods	CD-L2 ( $\text{mm}^{2}$ ) $\downarrow$			Fidelity Distance ( $\text{mm}^{2}$ ) $\downarrow$			Hausdorff Distance (mm) $\downarrow$			F-score $\uparrow$
Methods	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall	Premolar	Molar	Overall
MADCrowner (Point2Mesh)	0.247	0.238	0.240	0.129	0.145	0.140	1.841	1.832	1.834	0.842	0.849	0.847
MADCrowner (NKSR)	0.235	0.227	0.229	0.132	0.126	0.128	1.568	1.485	1.508	0.857	0.863	0.861
Ours	0.198	0.180	0.185	0.083	0.087	0.086	1.096	1.026	1.046	0.907	0.920	0.917

Table 6: Ablation study on postprocessing and performance comparison of different crown generation methods in mesh . The best results are highlighted in blue .

Figure 13: Visualization of different mesh reconstruction method results

Methods	Medial Area Difference ( $\text{mm}^{2}$ ) $\downarrow$	Lateral Area Difference ( $\text{mm}^{2}$ ) $\downarrow$
PCN+SAP [ 47 ]	6.51	7.30
TopNet+SAP [ 35 ]	15.97	16.50
GRnet+SAP [ 41 ]	5.06	4.90
DMCv2 [ 12 ]	5.18	4.72
VBCD [ 40 ]	5.00	5.04
Diffusion SDF [ 4 ]	5.37	5.18

Table 7: Proximal Contact Area deviation of different methods. The best and the second-best results are highlighted in blue and green , respectively.

Figure 15: The point cloud result from three models: DMCv2,GRnet+SAP,MADCrowner. To facilitate clearer visualization of the noisy point cloud, the predicted point cloud is downsampled to 2048 points using the farthest point sampler.

Figure 16: Side view of dental crowns. (a) The generated crown of the previous method, such as DMC. (b) The generated crown of MADCrowner without post-processing. (c) The generated crown of MADCrowner. (d) Ground Truth. The inherent watertight issue introduced by DPSR is addressed by our post-processing method.

Figure 17: Occlusal analysis. (a) Input IOS, (b) Generated crown of VBCD (c) Generated crown of MADCrowner, (d) Ground truth. For each vertex in the crown, the signed nearest distance to the IOS is visualized with a color map.

Figure 18: Detailed structures of a Geometry Aware Transformer (GAT), a Self Attention Transformer (SAT), and a Cross Attention Transformer (CAT)

Figure 19: Details of the Coarse Crown Refinement module. The number of points in the generated crown doubled after the processing.

Tooth Position	Number of Cases
	Maxillary	Mandible	Total
Incisor	217	33	250
Canine	115	13	128
Premolar	955	330	1285
Molar	1366	1951	3317

Table 8: Data distribution of the raw dataset for dental crown design.

Figure 20: Visualization for the generated dental crown on incisors and canines. (a) Input IOS mesh. (b) Generation results of the retrained model. (c) Ground truth.

Tooth Type	CD-L2 $\downarrow$	Fidelity Distance $\uparrow$	F-value $\uparrow$	Hausdorff Distance (mm) $\downarrow$
Incisor	0.285	0.143	0.766	1.589
Canine	0.406	0.174	0.731	1.820
Premolar	0.199	0.112	0.881	1.243
Molar	0.204	0.106	0.915	1.112

Table 9: Quantitative evaluation for dental crown generation across different tooth types.

3D Mesh Reconstruction

3D mesh reconstruction turns images, depth, or sensor data into explicit surfaces for measurement, editing, and simulation.

See all 109 papers

From Full and Partial Intraoral Scans to Crown Proposal: A Classification-Guided Restoration Assistance Pipeline
May 14, 2026Rabin Kunwar, Dikshya Parajuli, Rujal Acharya +123D Point Cloud Segmentation Dental Imaging
Accelerated Komputing Pvt. Ltd., Bhaktapur, Nepal. · University of Pittsburgh, Pittsburgh, PA, USA. · Emium Co. Ltd., Tokyo, Japan. +3
From Synthetic Data to Real Restorations: Diffusion Model for Patient-specific Dental Crown Completion
Mar 27, 2026Dávid Pukanec, Tibor Kubík, Michal Španěl3D Shape Completion Healthcare
Department of Computer Graphics and Multimedia, Brno University of Technology, Czechia
Deep Spectral Models for Robust Dental Shape Generation
Jun 30, 2026Tibor Kubík, François Guibault, Michal Španěl +1Geometric Representation Learning Generative Modeling
Polytechnique Montr´eal, Montr´eal, Canada · Brno University of Technology, Brno, Czech Republic

May 14, 2026J/K move · Enter open · S save

From Full and Partial Intraoral Scans to Crown Proposal: A Classification-Guided Restoration Assistance Pipeline

Rabin Kunwar, Dikshya Parajuli, Rujal Acharya, Romik Gosai+11

Accelerated Komputing Pvt. Ltd., Bhaktapur, Nepal. · University of Pittsburgh, Pittsburgh, PA, USA. · Emium Co. Ltd., Tokyo, Japan.+3

3D Point Cloud Segmentation Dental Imaging

Open paper AskPDF arXiv

Single-unit crown restoration is among the most common procedures in clinical dentistry, with CAD/CAM workflows now designing crowns directly from intraoral scans. Partial scans are often preferred over full-arch scans for single-unit cases due to fewer stitching errors, yet most segmentation networks trained on full arches fail on partial scans, while end-to-end generative crown methods often produce over-smoothed surfaces that lose occlusal detail. We propose an end-to-end pipeline that takes a raw intraoral scan and target FDI tooth number as input and outputs an initial, patient-specific crown proposal for clinician refinement. The pipeline has three phases: (I) data preparation and pose standardization; (II) segmentation routed by scan type; and (III) crown proposal generation via context-aware retrieval and Blender-based fitting. We address partial-scan segmentation through a classify-then-align strategy: a DGCNN classifier categorizes the scan into one of five anatomical types, then coarse-to-fine RANSAC+ICP registration standardizes the jaw coordinate frame, followed by graph-cut optimization to refine tooth-gingival boundaries. Trained on 1,958 partial scans, the pipeline achieves macro-average DSC 0.9249, Recall 0.8919, and Precision 0.9615 across 17 semantic classes; a fine-tuned full-arch model reaches DSC 0.9347. The prepared tooth and its mesial and distal neighbors achieve DSC 0.9468-0.9569 with sub-millimeter Centroid Errors (0.2666-0.2774 mm). These centroids anchor a retrieval module using DGCNN embeddings and cosine similarity over neighboring and opposing teeth, followed by spline-guided alignment and Blender Python API refinement. The pipeline produces a preliminary crown shell in 2.5-3.5 minutes, offering a practical alternative to end-to-end generative approaches.

Abstract

3D Mesh Reconstruction

Explore similar work

From Full and Partial Intraoral Scans to Crown Proposal: A Classification-Guided Restoration Assistance Pipeline

From Synthetic Data to Real Restorations: Diffusion Model for Patient-specific Dental Crown Completion

Deep Spectral Models for Robust Dental Shape Generation