A High-Fidelity Framework for 3D Dental Crown Reconstruction Using Intraoral Images

Highlights

• •A framework was developed for 3D crown reconstruction from five intraoral images.
• •The precision of reconstructed dental crowns improved via 3D Gaussian neural fields.
• •The reconstructed crown was precisely registered to the IOS model.
• •DenGaussDiff achieved robust performance on the public dataset of oral diseases.

Abstract

Introduction and aims

This study aimed to develop a novel framework, DenGaussDiff, which employed three-dimensional (3D) Gaussian neural fields and diffusion priors for precise dental crown reconstruction from 5 intraoral images.

Methods

This study collected 1000 clinical cases with 5 intraoral images and paired intraoral scanning (IOS) models and a public dataset was employed. First, images were segmented and enhanced to improve input quality. Next, 3D basic representation of the crowns was constructed using integrated camera pose estimation and an optimised ControlNet++ model. Then, a learnable neural representation was introduced and multiscale Gaussian rendering was applied to optimise the 3D crown representation. Finally, the Iterative Closest Point (ICP) algorithm was used for registration with IOS models.

Results

On our dental crown dataset, DenGaussDiff outperformed other reconstruction methods, with a 1.33% increase in Peak Signal-to-Noise Ratio (PSNR), a 3.9% increase in Structural Similarity Index (SSIM) and a 4.3% reduction in Learned Perceptual Image Patch Similarity (LPIPS). The point cloud registration results demonstrated a high-precision match between the reconstructed 3D crown model and IOS model. On the public dataset, DenGaussDiff achieved excellent performance among various evaluation indicators in dental plaque, tooth absence and tooth decay. Each key component of our framework was shown to be critical for precise reconstruction in ablation study.

Conclusion

The DenGaussDiff framework provides precise 3D dental crown reconstructions from sparse intraoral images, suggesting its potential in orthodontic diagnosis and monitoring. This further facilitates the clinical implementation of digital dentistry.

Introduction

With advances in artificial intelligence (AI) and its automatic deep learning processes, AI in dentistry is undergoing a significant revolution, enabling many applications such as diagnostics, treatment planning, patient education and imaging analysis.¹ Orthodontics, as a discipline which relies heavily on accurate images and models, particularly benefits from this revolution. Currently, AI has shown great potential to help orthodontists diagnose accurately and monitor treatment progress effectively.² For example, an AI system was designed for dental occlusion classification using intraoral images.³ Furthermore, other studies on AI applications have developed a fully automated deep learning model for tooth segmentation⁴ and a 3D dental model dataset was designed to enable automatic tooth alignment.⁵

Orthodontic treatment is a process that requires regular monitoring and adjustments over an extended period.⁶ For many years, intraoral photography has played a vital role in maintaining records and tracking treatment progress.^7,8 However, these images are unable to reflect the 3D spatial relationships among the teeth. Consequently, 3D dental crown models have become widely adopted by orthodontists for their clear visualization of dental structures. Currently, intraoral scanners are the primary technology for obtaining high-precision 3D digital dental crown models.⁹ Despite their effectiveness, factors such as high cost and the need for frequent in-person visits limit their widespread application. This limitation has driven AI-aided research toward automatic 3D dental crown model reconstruction from easily accessible 2D intraoral image data. Such research not only enhances doctor-patient communication but also enables precise monitoring of tooth movement, facilitating remote consultations and thereby improving the efficiency and flexibility of diagnosis and treatment.

In clinical photography during orthodontic treatment, at least 5 intraoral images were required according to the standardised guidelines: a frontal view, 2 lateral views and 2 occlusal views.¹⁰ These images are insufficient for dense input requirements, necessitating sparse-view 3D reconstruction. However, existing 3D reconstruction methods often struggle with insufficient accuracy when dealing with sparse view inputs.¹¹ Some methods require dense image inputs and complex device configurations, which are impractical in real-world orthodontic treatment scenarios.¹² While deep learning-based methods have made significant progress, there remains room for improvement in model generalisation and detail reconstruction.

In previous studies, researchers have proposed some methods to face the challenge with few images for 3D dental reconstruction. For instance, Chen et al.¹³ proposed a method for reconstructing 3D dental structures from 5 intraoral photographs using parameterised tooth models. However, the reconstruction quality largely depends on predefined tooth templates, making it difficult when facing complex clinical dental conditions. Liang et al.¹⁴ proposed a deep neural network framework to restore 3D teeth from a single panoramic X-ray. Song et al.¹⁵ reconstructed the 3D information of the entire oral cavity based on a single panoramic X-ray image and prior dental arch information. However, these methods still require patients to attend frequent follow-ups and expose patients to radiation. In summary, they are not suitable for remote monitoring in orthodontic treatment.

To address these challenges, we propose the DenGaussDiff framework, a complete process for reconstructing dental crowns from 5 intraoral images based on a 3D Gaussian neural field. First, intraoral images were collected and preprocessed as input data by the SAM model¹⁶ and the Otsu algorithm.¹⁷ Second, a 3D basic representation of dental crowns was reconstructed using a ControlNet++ model¹⁸ combined with camera pose estimation. Third, learnable neural representations were introduced to distill geometric details of diverse dental crowns and a multiscale Gaussian rendering strategy was integrated to achieve high-precision dental crown model. Finally, reconstructed crowns were registered with the paired IOS models to ensure their fidelity through the optimised ICP algorithm.^19,20 This framework enables precise reconstruction of 3D dental crown models, which makes it possible for orthodontic diagnosis and treatment and remote dental monitoring, effectively advancing the practical application of digital orthodontic diagnosis and treatment (Figure 1).

Fig. 1.

The framework of DenGaussDiff. This framework provides precise 3D dental crown reconstructions from sparse intraoral images using 3D Gaussian neural fields and diffusion priors.

Materials and methods

Source of data

The inclusion criteria were as follows: (1) male or female patients aged 10-40 years; (2) complete permanent dentition excluding third molars; (3) in good general health without systemic diseases affecting tooth development; (4) availability of high-quality intraoral images and paired IOS digital dental models. The exclusion criteria were as follows: (1) patients undergoing fixed orthodontic treatment; (2) severe enamel hypoplasia; (3) severe active periodontal disease; (4) multiple defective dental restorations in the oral cavity.

This study collected 1000 cases from the Orthodontics Department of the Affiliated Stomatological Hospital of Guangxi Medical University. Children and adolescents aged 10 to 18 accounted for about 70%, adults aged 19 to 35 accounted for about 20% and adults over 35 accounted for about 10%. The male-to-female ratio was 1: 2.18. Tooth morphology distribution was distributed as follows: incisors 28.3%, canines 14.2%, premolars 28.6% and molars 28.9%. The severity of dental crowding was approximately 40.2% mild, 11.6% moderate and 3.7% severe. Among all included patients, the distribution of Angle’s classifications was as follows: Class I (n = 493), Class II (n = 402) and Class III (n = 105). In addition, a total of 2.9% of patients had existing restorations.

Each case included 5 intraoral images and a paired IOS 3D model (Figure 2). All images were collected by nurses with standardised training using high-definition Nikon D7200 DSLR cameras with a macro lens. The aperture was set to f/18-22 with a shutter speed of 1/125 s and ISO of 200 to 250 and an auxiliary fill light was used as needed. For the frontal and lateral views, the occlusal plane was located in the middle of the bite images and the camera lens was perpendicular to the tooth surface. While in the maxillary and mandibular views, dental midline was kept cantered. The paired IOS models were obtained by iTero intelligent oral scanner. We used Cloudcompare software to manually extract the crown area of the scanned 3D model of the oral cavity as the basis for subsequent experiments. The 1000 cases were randomly divided into 3 subsets without considering age in the allocation. For model training, 800 cases were allocated for training, 100 for validation and 100 for testing.

Fig. 2.

A schematic diagram of dental crown data for each patient, including 5 intraoral images and the paired IOS models.

This study employed another public oral disease dataset consisting of dental condition images collected from multiple hospitals and well-known dental websites.²¹ This dataset includes 8188 2D natural images covering various dental conditions such as gingivitis, tooth discolouration and edentulism, which have been enhanced with data augmentation techniques such as rotation, flipping, scaling and noise addition. Unlike the dental crown dataset, the images in the Oral Diseases dataset lack sufficient images taken from multiple angles for 1 patient. Therefore, after careful screening, we selected 3 images that can describe the same patient's dental crown from 3 common dental diseases (dental plaque, tooth absence and tooth decay) as input. In this study, the image size of the dataset was adjusted to 640 × 640 pixels.

Preprocessing of the intraoral images

To optimise the input data quality, we preprocessed the intraoral images, retaining only the dental crown area in the image as the input image for model validation and comparison of 3D reconstruction effects. The Segment Anything Model (SAM)¹⁶ was employed to segment the dental crown area in the images inside the exit and obtain its mask image and the Otsu algorithm¹⁷ was employed to remove outliers and noise. In addition, the Augmentor and Imgaug image libraries were used to amplify dental crown image data.²² Considering interference factors present in clinical intraoral images, the enhancement techniques included brightness adjustment, adding noise and cropping.

3D basic representation of dental crowns

At first, a diffusion prior was introduced through ControlNet++, an image generation network built upon the stable diffusion model.¹⁸ Starting from the official open-source packages, we take the preprocessed dental crown image as the control input and adapt the conditional ControlNet++ input to predict RGB images, normal maps and depth maps separately in a single forward pass. These multimodal outputs are then used as geometric prior information for 3D Gaussian representation of the basic dental crown. Then, the idea of the RelPose++ model²³ was adopted to calculate the camera pose of n images input into the network, Y = {I₁, ..., I_n}, I_i ∈ R^{H × W × 3}. This is because in clinical photography, the number of available views of each patient is limited, leading to insufficient matching features for reconstruction. In addition, the 3D Gaussian representation incorporated optimisable texture parameters, enabling precise texture and illumination decoupling of the dental crown. The combination of these approaches facilitates the basic representation of dental crowns.

Optimisation of 3D dental crowns

To optimise the representation of dental crowns, a learnable neural representation module was proposed, which exploits neural fields to capture accurate colour and texture information of the crown. In the architecture of learnable neural representation, geometric structure and appearance are first encoded by a compact neural field rather than directly by the sparse 3D Gaussians. A neural rendering network with multiresolution hash encoding is used to represent the dental crown surface as a continuous implicit function in 3D space. On top of this implicit geometry, 1 branch predicts a view-independent diffuse colour and a weight controlling the strength of specular effects, while another branch takes the viewing direction as input and produces view-dependent specular lighting. The network is optimised with an image reconstruction loss together with an additional regularisation term on the implicit surface function and the optimised normal and depth values are then used to guide the densification and refinement of the 3D Gaussian distribution.

By means of this module, initial normal maps and depth maps were distilled, while the geometric information was refined due to diversified 3D Gaussian elements. A dense grid system was established based on the learnable neural representation, which can accurately query volume data of various resolutions and effectively capture detailed surface geometry even in sparsely sampled areas. We identified the key surface transition areas of the grid by setting a density threshold and interpolate the surface intersections, which provides dense reconstruction and geometric perception of basic 3D Gaussian representations.

Then a multiscale Gaussian rendering strategy was introduced. For each Gaussian element, we first evaluate its contribution in all input views by accumulating the opacity inside its 2D projection footprint and choose the view with the highest contribution as a reference. From this reference view, depth and surface normal for the Gaussian are estimated from the neural field and then checked against the geometry predicted at other views, so that inconsistent or unreliable estimates can be suppressed. We further compute a relative difference between these refined values and the initialisation, which allows us to separate slowly varying areas from sharp and high-frequency regions. Regions with small differences are directly back to update the 3D crown reconstruction, whereas regions with large differences are re-rendered for additional refinement. This multiscale filtering and update process alleviates aliasing caused by small Gaussian elements and improves both geometric fidelity and visual quality of the reconstructed dental crowns. By selecting the view with the highest contribution, calculating the filtered depth and normal values and controlling high-frequency regions, the strategy improved accuracy and quality in dental crown 3D reconstruction (Figure 3).

Fig. 3.

Flowchart of 3D dental crown optimisation.

Registration between reconstructed crowns and IOS models

We validated the reconstructed crowns by registering them with the paired IOS models. ICP and optimised ICP algorithm were used to compare the homologous points between the 2 models.^19,20 The IOS models were marked with green point clouds, while reconstructed crowns were marked with red point clouds. During the registration of the green and red point clouds, each noncoinciding point was computed as the registration error and expressed in mm. Lower registration error indicates greater spatial similarity with the IOS models.

Evaluation metrics and statistical analysis

In all experiments, the model was optimised for 10,000 iterations with a batch size of 2. The learning rate was scheduled to linearly grow from e^-5 at the start of training to 5e^-4 at the end. To quantitatively verify the advantages of this framework, we used 3 widely used indicators to evaluate, namely, PSNR, SSIM and LPIPS.²⁴ To evaluate the accuracy of the reconstructed 3D model of dental crowns, we calculated chamfer distance (CD), the F1 score and the Intersection over Union index (IoU) under different thresholds according to the reference mesh.²⁵ All statistical analyses were performed using GraphPad Prism software. Paired t-tests were used to compare the registration error of DenGaussDiff and other methods within the same patient, P <.05 was considered significant.

Results

Qualitative and quantitative results

For sparse view input, we selected existing advanced methods in the field of computer vision, including FSGS (Zhu et al.),²⁶ SAP3D (Han et al.),²⁷ MVSplat360 (Chen et al.),²⁸ and TeethDreamer (Xu et al.),²⁹ to compare with DenGaussDiff. The visualisation results of various advanced sparse view 3D reconstruction methods on the dental crown dataset were shown in Figure 3 and their quantitative indicators were shown in Table 1. The DenGaussDiff framework proposed in this article was more in line with the real dental crowns in the reconstruction results and can construct dental crown meshes containing fine textures (Figure 4). Among various evaluation indicators, the DenGaussDiff framework has shown excellent performance. In the reconstruction of maxillary dental crowns, PSNR achieved 18.58, SSIM achieved 0.827, LPIPS achieved 0.153, CD achieved 0.208, F1 score achieved the highest value of 0.807 and IOU achieved 0.405. In the reconstruction of the mandibular dental crown, PSNR achieved 16.57, SSIM achieved 0.747, LPIPS achieved 0.203, CD achieved 0.244 and the F1 score achieved the highest value of 0.787, while IOU achieved 0.359 (Table 1).

Table 1.

Quantitative comparison of the DenGaussDiff framework with other methods on the dental crown dataset.

Region	Method	PSNR↑	SSIM↑	LPIPS↓	CD(mm)↓	F1@0.04↑	IoU↑
Maxillary	SAP3D	16.75	0.748	0.278	0.441	0.572	0.142
	FSGS	16.89	0.742	0.205	0.325	0.615	0.251
	MVSplat360	17.07	0.783	0.147	0.295	0.738	0.341
	TeethDreamer	17.25	0.804	0.176	0.242	0.772	0.376
	DenGaussDiff	18.58	0.827	0.153	0.208	0.807	0.405
Mandibular	SAP3D	14.28	0.518	0.411	0.445	0.575	0.192
	FSGS	14.97	0.589	0.348	0.381	0.617	0.241
	MVSplat360	15.45	0.642	0.304	0.345	0.688	0.287
	TeethDreamer	16.15	0.708	0.246	0.294	0.725	0.324
	DenGaussDiff	16.57	0.747	0.203	0.244	0.787	0.359

Fig. 4.

Qualitative comparison of the DenGaussDiff framework with other methods on the dental crown dataset.

Registration between reconstructed crowns and IOS models

To explore the potential for clinical application of DenGaussDiff, we compared its reconstructed crowns with the clinical gold standard, IOS models, by 3D point cloud registration. Results showed high spatial consistency between reconstructed crowns and IOS models in both the maxillary and mandibular regions (Figures 5A and B). Visual comparisons at each tooth site provided further insights (Figures 5C and D). For the same tooth in the same patient, paired t-tests showed that the registration errors of DenGaussDiff with the IOS model were significantly lower than those of other methods (Figures 5E and F). Table 2 further reported the associated statistical metrics, with low P-values and large Cohen’s d supporting this difference.

Fig. 5.

3D Registration between reconstructed crowns and IOS Models. (A) Point cloud registration of maxillary models between DenGaussDiff and IOS. (B) Point cloud registration of mandibular models between DenGaussDiff and IOS. (C) Visualisation of registration accuracy for maxillary teeth. (D) Visualisation of registration accuracy for mandibular teeth. (E) Maxillary mean registration error of different methods. (F) Mandibular mean registration error of different methods. ***P <.001.

Table 2.

Statistical analysis of registration error among different methods using paired-t tests.

	Paired-t test	P-value	95% confidence intervals	Cohen's d
Maxillary	SAP3D-DenGaussDiff	<.001	0.577 to 0.604	7.27
	FSGS-DenGaussDiff	<.001	0.549 to 0.585	5.34
	MVSplat360-DenGaussDiff	<.001	0.413 to 0.442	4.90
	TeethDreamer-DenGaussDiff	<.001	0.349 to 0.377	4.40
Mandibular	SAP3D-DenGaussDiff	<.001	0.580 to 0.610	6.79
	FSGS-DenGaussDiff	<.001	0.532 to 0.565	5.57
	MVSplat360-DenGaussDiff	<.001	0.388 to 0.411	5.67
	TeethDreamer-DenGaussDiff	<.001	0.312 to 0.332	5.36

Evaluation of DenGaussDiff on the public dataset

The visualisation results of various advanced sparse view 3D reconstruction methods on the Oral Diseases dataset were shown in Figure 6 and their quantitative indicators were shown in Table 2. In the visualisation results, the DenGaussDiff framework proposed in this article was closer to the real clinical conditions, indicating better morphological and detail fidelity. It can render the colour of dental plaque and the situation of tooth decay (Figure 6). Quantitative comparison showed that in the presence of dental plaque, tooth absence and tooth decay, DenGaussDiff still achieved excellent performance across various evaluation indicators (Table 3).

Fig. 6.

Qualitative comparison of the DenGaussDiff framework with other methods on the Oral Diseases dataset.

Table 3.

Quantitative comparison of the DenGaussDiff framework with other methods on the Oral Diseases dataset.

Region	Method	PSNR↑	SSIM↑	LPIPS↓	CD (mm)↓		F1@0.04↑	IoU↑
Dental plaque	SAP3D	11.89	0.515	0.511	0.591		0.469	0.139
	FSGS	13.93	0.627	0.385	0.386		0.601	0.229
	MVSplat360	14.04	0.647	0.371	0.356		0.605	0.266
	TeethDreamer	14.99	0.715	0.271	0.288		0.682	0.302
	DenGaussDiff	15.53	0.721	0.258	0.261		0.719	0.334
Tooth absence	SAP3D	14.77	0.545	0.482	0.528		0.579	0.188
	FSGS	17.10	0.698	0.265	0.259		0.762	0.328
	MVSplat360	16.50	0.668	0.312	0.325		0.678	0.311
	TeethDreamer	17.46	0.714	0.241	0.239		0.776	0.332
	DenGaussDiff	17.71	0.755	0.198	0.236		0.791	0.385
Tooth decay	SAP3D	12.47	0.541	0.442	0.523		0.598	0.192
	FSGS	13.54	0.604	0.355	0.405		0.689	0.225
	MVSplat360	14.73	0.678	0.272	0.343		0.748	0.334
	TeethDreamer	14.67	0.648	0.304	0.381		0.711	0.276
	DenGaussDiff	15.29	0.702	0.251	0.304		0.762	0.367

Ablation study

Ablation studies were designed to validate the effectiveness of the key components of DenGaussDiff at each stage. Three models under the same conditions as the DenGaussDiff framework were trained to analyse the effectiveness of each component. In the 3D basic representation of dental crowns stage, we constructed a DenGaussDiff w/o A1 model that removed the computational camera pose and ControlNet++Stable Diffusion model from the DenGaussDiff framework. It can be intuitively seen that adding camera pose calculation and ControlNet++ stable diffusion model can significantly improve the geometric accuracy and visual effect of 3D dental crown reconstruction, thereby providing more stable data support for downstream tasks of dental crown 3D reconstruction (Figure 7). The DenGaussDiff w/o A1 model was 1.13 lower than the DenGaussDiff model in PSNR index, 0.059 lower in SSIM index and 0.05 higher in LPIPS index (Table 4).

Fig. 7.

Visualisation effect of DenGaussDiff framework in 2-stage ablation experiments.

Table 4.

Results of the ablation experiments for the key components of the DenGaussDiff framework.

Method	PSNR↑	SSIM↑	LPIPS↓
DenGaussDiff w/o A3	13.89	0.521	0.387
DenGaussDiff w/o A1	14.68	0.589	0.289
DenGaussDiff w/o A2	14.55	0.547	0.311
DenGaussDiff	15.81	0.648	0.239

In the optimisation of 3D dental crowns phase, we constructed a DenGaussDiff w/o A2 model, which removed the learnable neural representation of the Optimised 3D Gaussian representation of dental crowns stage and the Multi scale Gaussian Rendering Strategy. The DenGaussDiff w/o A2 model and the DenGaussDiff model were compared while enlarging the details of the reconstructed dental crown 3D model. It can be intuitively seen that adding learnable neural representations and multiscale Gaussian rendering strategies can refine the geometric accuracy and rendering effect of the Gaussian representation of dental crowns (Figure 7). The DenGaussDiff w/o A2 model was 1.26 lower than the DenGaussDiff model in PSNR metrics, 0.101 lower in SSIM metrics and 0.072 higher in LPIPS metrics (Table 4).

Finally, a DenGaussDiff w/o A3 model, completely deleting the above content, was constructed. It was 1.92 lower than the DenGaussDiff model in PSNR metrics, 0.127 lower in SSIM metrics and 0.148 higher in LPIPS metrics (Table 4).

Discussion

Orthodontic diagnosis and treatment mainly rely on precise 3D crown models to assess tooth alignment, occlusal relationships and monitor progress.³⁰ While IOS offer high accuracy, their clinical utility is limited by cost and patient access – particularly for patients who struggle with frequent in-person visits.³¹ Conventional 2D photos fail to provide enough 3D details needed for monitoring. This limitation necessitates methods to reconstruct precise crown models from sparse, accessible intraoral images. To meet this need, we propose DenGaussDiff, a high-fidelity framework for dental crown reconstruction via 3D Gaussian neural fields and diffusion priors.

For input images, suitable imaging perspectives are required to cover all anatomical crown surfaces. Clinically, at least 5 intraoral views were taken for each patient: a frontal view with teeth in occlusion, left/right buccal views and 2 occlusal views.¹⁰ Such views can also be easily learned and reproduced by patients for self-photography.³² In addition to imaging perspectives, high and consistent-resolution images are recommended to achieve more accurate reconstructed models in our methods. Currently, the resolution of smartphones has improved rapidly and is sufficient to meet the input requirements.³³ For future home use, such standardisation could be further supported by providing patients with simple cheek retractors and clear step-by-step instructions delivered through a mobile application on how to position the camera and open the mouth so that all teeth are visible. In addition, similar smartphone applications have been reported that help patients capture standardised intraoral photographs with sufficient accuracy for clinical use.³³ As a result, reconstruction can be achieved using these images obtained at the patient’s home through our framework, significantly advancing the application of digital orthodontics.

AI-aided 3D reconstruction is of great importance in clinical application,³⁴ so some key methods in 3D Gaussian neural fields were incorporated in DenGaussDiff framework. Firstly, SAM, as a basic segmentation model, provides reliable crown data for subsequent reconstruction. Then, ControlNet++ with camera pose calculation can generate multiple views from patients’ the intraoral images and ensure geometric consistency among the views.¹⁸ On this basis, learnable neural representations were used to optimise dental crown reconstruction from multiple perspectives through a dense grid system and further enhanced rendering efficiency combining multiscale Gaussian rendering.³⁵ The results of ablation studies also showed that the synergistic design of this technical route significantly enhanced the fidelity and consistency of the dental crown structure. Therefore, high-fidelity dental crown reconstruction was achieved using only 5 intraoral images, thereby facilitating clinical adoption including remote diagnosis and monitoring based on the DenGaussDiff framework.

On our dental crown dataset, qualitative and quantitative evaluations against existing methods showed DenGaussDiff's superiority across key indicators including PSNR, SSIM, LPIPS, F1-score and IoU. Improvement in PSNR and SSIM indicates better retention of structural information,³⁶ which is especially important for the identification of crown boundaries, occlusal and interdental contact points in orthodontics. A lower LPIPS reflects great microstructural reconstruction,³⁷ making DenGaussDiff capable of detecting early enamel demineralisation. Higher F1-score and IoU indicates more stability of dental segmentation and recognition,³⁸ supporting subsequent digital measurements like tooth position analysis, space analysis, the curve of Spee, length and width evaluation. Thus, precise dental crown reconstruction enables DenGaussDiff to be applied in the diagnosis of diverse malocclusions clinically.

Moreover, DenGaussDiff achieved crown structural preservation with a maxillary CD of 0.208 mm and a mandibular CD of 0.244 mm. A lower CD indicates that the geometric position and shape are closer to the actual structure,³⁹ enabling its application in orthodontic monitoring. This is because orthodontic monitoring relies on the precise detection of small tooth displacements. For example, during clear aligner treatment, tooth movement are typically planned in steps of 0.25 to 0.33 mm per aligner.⁴⁰ However, multiple clinical studies have reported that actual tooth movement may differ from the predesigned plan.^41,42 Possible reason is that some factors can cause variations in the rate of tooth movement like vertical skeletal pattern, patient compliance and age, which in turn leads to a mismatch between dental crown and clear aligners.^43,44 With our precisely reconstructed models during orthodontic treatment, this mismatch can be identified early, allowing timely adjustment of clear aligners.

In addition, a high proportion of point overlap showed that reconstructed crowns maintained good geometric consistency with the paired IOS model. Although DenGaussDiff still requires more clinical studies, there is potential for supporting computer-aided design and fabrication of appliances like clear aligners in the future. This advancement is expected to make clinical treatment more convenient for both patients and orthodontists.

Clinically, orthodontic patients often present with complex oral health conditions.⁴⁵ Therefore, DenGaussDiff was evaluated on the public dataset for its ability to capture pathological details in the presence of dental plaque, tooth absence and tooth decay. According to qualitative assessment, the reconstruction quality has nearly reached the standards for visual clinical diagnosis (Figure 6). For crowns with dental plaque, it achieved an F1 score of 0.719, suggesting that it can effectively capture the fine textural details and colour distribution in plaque. In cases with tooth absence, DenGaussDiff reflects its capability to capture the spatial relationships between the missing tooth and adjacent teeth. In addition, its precise rendering of carious regions, including changes in colour and texture, indicates potential in supporting visual-based early caries ICDAS scoring.

In orthodontic clinical research, measurement discrepancies of less than 0.5 mm between digital and plaster models have been reported as clinically insignificant.⁴⁶ Another clinical study suggested that discrepancies of 0.35 mm and 0.4 mm are clinically acceptable for specific orthodontic measurements.⁴⁷ On our orthodontic crown dataset, the CD values were 0.208 mm and 0.244 mm and on the public datasets the values were 0.261 mm, 0.236 mm and 0.304 mm. All these results fall well below the clinical tolerance threshold.

In summary, the potential clinical value of the DenGaussDiff framework is anticipated to be applied in the following 3 aspects: (1) Digital epidemiological surveys: In our framework, patients are only required to provide 5 intraoral images that are easily obtained, from which precise dental models can be reconstructed. This enables digital epidemiological analyses of malocclusion and other oral diseases, which is especially beneficial for patients who have limited access to regular oral examinations. (2) Early caries detection: In response to the increased risk of plaque retention caused by orthodontic appliances, DenGaussDiff can detect microstructural changes at the early stage of enamel demineralisation, enabling early caries detection. (3) Remote orthodontic monitoring: In long-term orthodontic treatments such as clear aligner therapy, sequential dental models can be reconstructed at the different stages of treatment. By analysing these reconstructions, clinicians can accurately assess tooth movement alignment and promptly determine whether intervention is needed, significantly reducing the number of patient visits.

This study also has certain limitations. Foremost, the DenGaussDiff framework is restricted to crown morphology reconstruction and cannot reconstruct complete dental arch or gingival anatomy. Our framework will be further developed to reconstruct complete dental arch, allowing more detailed information for clinicians. Additionally, the current validation of the reconstructed crown models mainly relies on quantitative and qualitative experiments and alignment with paired IOS models. Future studies are required to evaluate the application of DenGaussDiff in clinical practice. Recent reports also emphasise that dental AI still faces risks and challenges, indicating the need for improved patient communication and responsible implementation of our framework.⁴⁸

Conclusions

Our DenGaussDiff framework was highly accurate in 3D dental crown reconstruction using 5 intraoral images. It achieved significant improvements over other methods across different indicators in qualitative and quantitative experiments. In contrast with the clinical intraoral scanning method, it shows high agreement with the IOS model. Moreover, DenGaussDiff demonstrated the potential to detect common oral diseases, including dental plaque, tooth absence and tooth decay, in the public dataset. Further work is needed to bring the DenGaussDiff framework to clinical practice, contributing to the advancement of digital orthodontics.

Author contributions

Acquisition of data: Ni and Xinyu；analysis, or interpretation of data: Wenkang; Drafting of the manuscript: Ni and Xinyu; Critical revision of the manuscript for important intellectual content: Ni, Xinyu, Xuejun and Shuixue; Supervision: Xuejun and Shuixue.

Funding

This work was supported in part by the National Natural Science Foundation of China [grant numbers 82360182, 82260197], in part by the Guangxi Medical and Health Appropriate Technology Development and Promotion Application Project under Grant S2023095, in part by the Clinical Key Specialty Construction Project of the Affiliated Stomatological Hospital of Guangxi Medical University under Grant LCZDZK2023015, in part by the Guangxi Natural Science Foundation [grant numbers 2025GXNSFAA069789].

Conflicts of interest

None disclosed.