PX2Tooth: Reconstructing the 3D Point Cloud Teeth from a Single Panoramic X-ray

TL;DR

PX2Tooth, a novel approach to reconstruct 3D teeth using a single PX image with a two-stage framework, is proposed and a novel tooth generation network (TGNet) that learns to transform random point clouds into 3D teeth is designed.

Abstract

Reconstructing the 3D anatomical structures of the oral cavity, which originally reside in the cone-beam CT (CBCT), from a single 2D Panoramic X-ray(PX) remains a critical yet challenging task, as it can effectively reduce radiation risks and treatment costs during the diagnostic in digital dentistry. However, current methods are either error-prone or only trained/evaluated on small-scale datasets (less than 50 cases), resulting in compromised trustworthiness. In this paper, we propose PX2Tooth, a novel approach to reconstruct 3D teeth using a single PX image with a two-stage framework. First, we design the PXSegNet to segment the permanent teeth from the PX images, providing clear positional, morphological, and categorical information for each tooth. Subsequently, we design a novel tooth generation network (TGNet) that learns to transform random point clouds into 3D teeth. TGNet integrates the segmented patch information and introduces a Prior Fusion Module (PFM) to enhance the generation quality, especially in the root apex region. Moreover, we construct a dataset comprising 499 pairs of CBCT and Panoramic X-rays. Extensive experiments demonstrate that PX2Tooth can achieve an Intersection over Union (IoU) of 0.793, significantly surpassing previous methods, underscoring the great potential of artificial intelligence in digital dentistry.

Figures (4)

• Figure 1: Comparison of results from different 3D generation methods. Compared to other 3D generation methods, PX2Tooth (ours) can generate more detailed meshes representing tooth shapes, with significant improvements in the 3D representation of the root tip and the smoothness of the mesh surface.
• Figure 2: (a) The PX image is segmented into 32 teeth classes using the TeethNet model. Subsequently, a segmented patch from each teeth is generated by cropping the input PX image with the predicted segmentation mask, which incorporated into the initial point cloud using a new strategy. (b) In the reconstruction process, each point undergoes processing through a symmetry function. Additionally, every point in the point cloud is mapped to a low-dimensional feature space, enabling the capture of global information. The neural network structure is instrumental in mapping each point and its neighborhood, facilitating local feature learning to glean pertinent local structure information. In addition, to address the spatial misalignment issue, we suggest employing the Prior Fusion Module to seamlessly incorporate the spatial position information derived from the segmented patches into the generation process.
• Figure 3: Visual representation of sample outputs. A represents the comparison with the results of other methods, and B represents the comparison results of the ablation experiment.
• Figure 4: FDI shows the position of each tooth corresponding to the oral cavity, and Part A shows the visualization problem of teeth that are not satisfactory enough (lack of root tip information). Part B shows the category tooth visualization with excellent average high IoU performance.