X-Splat: Gaussian Splatting for 3D CBCT Generation from Single Panoramic Radiograph

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Imaging & 3D Reconstruction · Depth: Expert, quick

Summary

X-Splat is a new Gaussian Splatting framework designed to generate 3D dental volumes, similar to Cone-Beam Computed Tomography (CBCT), from a single panoramic radiograph (PXR). This addresses the challenge of reconstructing 3D anatomy from a 2D PXR, a highly underdetermined problem where existing implicit and generative methods often yield oversmoothed or inconsistent results. X-Splat leverages the known panoramic acquisition geometry, initializing learnable anisotropic Gaussian primitives along the X-ray paths of the input image. These primitives are adjusted in a single feed-forward pass, guided by Beer-Lambert reprojection and multi-view radiographic supervision. A residual refiner integrates dataset-level anatomical priors without compromising the Gaussian-resolved geometry. Trained on synthetic PXR-CBCT pairs, X-Splat enables direct volumetric supervision and introduces segmentation-based geometry-aware metrics for maxillofacial anatomy evaluation. It surpasses NeRF- and GAN-based baselines, accurately recovering individual teeth, cortical boundaries, alveolar structure, and the mandibular canal.

Key takeaway

For AI Scientists developing medical imaging solutions, X-Splat offers a robust approach to 3D reconstruction from sparse 2D inputs. You should consider integrating known acquisition geometry and Gaussian Splatting for improved anatomical detail and consistency, especially when dealing with underdetermined problems like PXR-to-CBCT conversion. This method demonstrates superior reconstruction of fine structures, reducing reliance on high-radiation CBCT scans. Explore synthetic data generation to enable direct volumetric supervision in similar contexts.

Key insights

X-Splat uses Gaussian Splatting and acquisition geometry to reconstruct accurate 3D dental volumes from single 2D panoramic radiographs.

Principles

• Underdetermined 3D reconstruction benefits from geometry-driven supervision.
• Synthetic data enables direct volumetric supervision for PXR-CBCT pairs.
• Combining Gaussian primitives with a residual refiner improves anatomical detail.

Method

X-Splat initializes anisotropic Gaussian primitives along known PXR X-ray paths, adjusting them via Beer-Lambert reprojection and multi-view supervision. A residual refiner then adds anatomical priors.

In practice

• Apply Gaussian Splatting for underdetermined 3D medical image reconstruction.
• Develop synthetic paired datasets for direct volumetric supervision.
• Integrate acquisition geometry as a scaffold for 3D generation.

Topics

• Gaussian Splatting
• 3D Reconstruction
• Cone-Beam CT
• Panoramic Radiography
• Medical Imaging
• Computer Vision

Code references

• tomek1911/X-Splat

Best for: AI Scientist, Research Scientist

Related on AIssential

• Three-dimensional Damage Visualization of Civil Structures via Gaussian Splatting-enabled Digital Twins — Gaussian Splatting enhances 3D damage visualization in civil infrastructure digital twins, improving accuracy and efficiency.
• Splaxel: Efficient Distributed Training of 3D Gaussian Splatting for Large-scale Scene Reconstruction via Pixel-level Communication — Splaxel's pixel-level communication for distributed 3DGS training drastically cuts overhead, enabling efficient scaling for large-scale scene reconstruction.
• MoonSplat: Monocular Online Gaussian Splatting with Sim(3) Global Optimization — MoonSplat enables robust, efficient online 3D reconstruction from monocular video using globally optimized voxelized 3D Gaussian Splatting.
• Deep Learning Approaches for 3D Medical Scene Completion: From Geometric Modeling to Generative Paradigms — The field of 3D medical scene completion has rapidly evolved from voxel-based methods to advanced generative and rendering-aware paradigms.
• Intrinsic decomposition and editing of 3D Gaussian splats — Intrinsic decomposition for 3D Gaussian splats enables texture editing independent of lighting across viewpoints.
• EnerGS: Energy-Based Gaussian Splatting with Partial Geometric Priors — EnerGS uses a soft energy field from partial geometry to guide 3D Gaussian Splatting, improving reconstruction quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.