3D Teeth Reconstruction from Panoramic Radiographs Using Neural Implicit Functions

Sihwa Park, Seongjun Kim, In-Seok Song, Seung Jun Baek

Abstract

Panoramic radiography is a widely used imaging modality in dental practice and research. However, it only provides flattened 2D images, which limits the detailed assessment of dental structures. In this paper, we propose Occudent, a framework for 3D teeth reconstruction from panoramic radiographs using neural implicit functions, which, to the best of our knowledge, is the first work to do so. For a given point in 3D space, the implicit function estimates whether the point is occupied by a tooth, and thus implicitly determines the boundaries of 3D tooth shapes. Firstly, Occudent applies multi-label segmentation to the input panoramic radiograph. Next, tooth shape embeddings as well as tooth class embeddings are generated from the segmentation outputs, which are fed to the reconstruction network. A novel module called Conditional eXcitation (CX) is proposed in order to effectively incorporate the combined shape and class embeddings into the implicit function. The performance of Occudent is evaluated using both quantitative and qualitative measures. Importantly, Occudent is trained and validated with actual panoramic radiographs as input, distinct from recent works which used synthesized images. Experiments demonstrate the superiority of Occudent over state-of-the-art methods.

Introduction

In modern dentistry, clinicians rely on two primary imaging types: Panoramic X-rays (PX), which provide a flattened 2D view of the entire mouth, and Cone-Beam Computed Tomography (CBCT), which provides a full 3D volume. While CBCT is superior for planning complex procedures like implants, it is significantly more expensive and exposes patients to much higher doses of radiation. Reconstructing 3D teeth from a 2D PX is a classic "ill-posed" problem because information is lost during the 2D projection. Occudent addresses the challenge of generating accurate 3D models from real-world 2D panoramic radiographs, which contain noise and artifacts not present in synthetic datasets.

The Occudent Framework

The Occudent framework operates in three main stages to bridge the gap between 2D observations and 3D reconstruction:

  1. Multi-label 2D Segmentation: The input PX image is processed by a UNet++ model to identify and isolate 32 individual tooth classes. This allows the system to treat each tooth as a distinct object.
  2. Condition Generation: For each identified tooth, the system creates a "condition vector" combining Class Embeddings (learned representations of specific tooth types) and Shape Embeddings (features extracted from a cropped 2D patch of the original X-ray using a ResNet-18 encoder).
  3. Neural Implicit Reconstruction: Instead of using memory-heavy voxels, Occudent uses a Neural Implicit Function. Given any 3D coordinate $(x, y, z)$, the network predicts the probability that the point is "occupied" by a tooth.
  4. Conditional eXcitation (CX): The authors introduce a novel CX module. Inspired by Squeeze-and-Excitation Networks, CX uses the condition vector to dynamically scale the features within the reconstruction network, injecting 2D observations into the 3D space more effectively than standard normalization techniques.
Overview of the Occudent framework, including segmentation, condition generation, and the CX-based reconstruction pipeline.

Experimental Results

Occudent was pre-trained on 4,000 PX images from "The Open AI Dataset Project" and validated using a specialized dataset of 39 real-world PX images paired with matched CBCT scans from Korea University Anam Hospital. Unlike many competitors that use synthetic projections, this study uses actual clinical radiographs.

Occudent outperformed several state-of-the-art baselines (including 3D-R2N2, PSGN, and OccNet) across all quantitative metrics:

  • IoU (Intersection over Union): Achieved 0.651, significantly higher than the next best baseline (OccNet at 0.611).
  • Chamfer-L1 Distance: Reduced error to 0.298, compared to 0.353 for OccNet.

Qualitatively, visual results show that Occudent is particularly effective at reconstructing complex molar roots, a feature that most other models either over-simplify or fail to generate entirely.

A high-fidelity 3D tooth model reconstructed by Occudent, demonstrating detailed root structure.

Conclusion

Occudent demonstrates the feasibility of high-fidelity 3D dental reconstruction from standard 2D radiographs using neural implicit functions and the novel CX module. By utilizing anatomical priors and effective feature scaling, the model produces continuous surfaces that avoid the resolution limits of voxel-based approaches. Future work will focus on expanding the clinical dataset and exploring multi-modal integration to further refine meshes for surgical planning.

Summary updated

Abstract

Panoramic radiography is a widely used imaging modality in dental practice and research. However, it only provides flattened 2D images, which limits the detailed assessment of dental structures. In this paper, we propose Occudent, a framework for 3D teeth reconstruction from panoramic radiographs using neural implicit functions, which, to the best of our knowledge, is the first work to do so. For a given point in 3D space, the implicit function estimates whether the point is occupied by a tooth, and thus implicitly determines the boundaries of 3D tooth shapes. Firstly, Occudent applies multi-label segmentation to the input panoramic radiograph. Next, tooth shape embeddings as well as tooth class embeddings are generated from the segmentation outputs, which are fed to the reconstruction network. A novel module called Conditional eXcitation (CX) is proposed in order to effectively incorporate the combined shape and class embeddings into the implicit function. The performance of Occudent is evaluated using both quantitative and qualitative measures. Importantly, Occudent is trained and validated with actual panoramic radiographs as input, distinct from recent works which used synthesized images. Experiments demonstrate the superiority of Occudent over state-of-the-art methods.
BibTeX
@inproceedings{park2023d,
  author = {Sihwa Park and Seongjun Kim and In-Seok Song and Seung Jun Baek},
  title = {{3D Teeth Reconstruction from Panoramic Radiographs Using Neural Implicit Functions}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=bfoD7K9FAB},
}
Figure