💻 computer science

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

X-Splat is a novel Gaussian Splatting framework that generates high-fidelity 3D dental CBCT volumes from a single panoramic radiograph by leveraging acquisition geometry and learnable anisotropic primitives to overcome the underdetermined nature of the problem, outperforming existing NeRF and GAN methods in reconstructing sharp anatomical structures like the mandibular canal.

Original authors: Tomasz Szczepański, Szymon Płotka, Michal K. Grzeszczyk, Tomasz Trzciński, Arkadiusz Sitek

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tomasz Szczepański, Szymon Płotka, Michal K. Grzeszczyk, Tomasz Trzciński, Arkadiusz Sitek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess the shape of a complex 3D sculpture, but you are only allowed to look at a single, flat shadow it casts on a wall. That is the challenge dentists face when they try to create a 3D model of a patient's jaw from a standard 2D panoramic X-ray. The X-ray squashes all the depth information into one flat image, leaving the "third dimension" a mystery.

Previous attempts to solve this were like trying to sculpt the jaw out of smooth, melted clay. They could get the general shape, but they missed the sharp edges, the tiny roots of teeth, and the delicate tunnels (like the mandibular canal) that run through the jawbone. They often "hallucinated" features that weren't there or smoothed over important details.

Enter X-Splat: The "Smart Confetti" Approach

The authors propose a new method called X-Splat. Instead of using smooth clay, they use thousands of tiny, invisible, 3D "confetti pieces" (mathematically known as Gaussian primitives). Here is how it works, using simple analogies:

1. The Scaffold: Following the Light Beams

Think of the panoramic X-ray machine as a flashlight rotating around a patient's head. The paper uses the known path of these light beams as a "scaffold" or a skeleton.

The Old Way: Imagine trying to build a house by guessing where the walls should go in empty space.
The X-Splat Way: Imagine placing a tiny, stretchy balloon exactly on every single path that the flashlight beam took. You have millions of these balloons, all lined up along the known paths of the X-rays.

2. The Magic: Stretching and Rotating

Once these balloons are placed, a smart computer program (a neural network) tells them how to change shape.

Stretching: If the X-ray beam passed through a thick tooth, the balloons along that path stretch out to fill the space.
Rotating: If the beam hit a curved root, the balloons twist to match that curve.
The Result: These balloons snap together to form sharp, precise boundaries for teeth and bone, rather than a blurry blob. They can even fill in the gaps between the light beams, creating a solid 3D volume.

3. The Safety Net: The "Residual Refiner"

Sometimes, just stretching balloons isn't enough to get every tiny detail perfect. The paper adds a second step: a "lightweight refiner."

Think of this as a master sculptor who looks at the balloon structure and makes tiny, final adjustments.
Crucially, this sculptor is very small and cautious. They are only allowed to make small tweaks based on what they know about average jaw shapes. They are not allowed to completely redesign the structure, ensuring the final result still matches the original X-ray perfectly.

4. The Training: Learning from "Fake" Scans

To teach this system, the researchers didn't use real patients (because getting a perfect 2D X-ray and a matching 3D scan of the same person at the exact same moment is nearly impossible).

Instead, they used a massive library of real 3D jaw scans.
They used a computer to simulate what the 2D X-ray would look like for each 3D scan.
They taught X-Splat to turn that simulated 2D shadow back into the original 3D shape. Because they had the "answer key" (the original 3D scan), the system learned exactly how to stretch and rotate those balloons to get it right.

Why It's Better (The Results)

The paper compares X-Splat to other methods (like NeRF and GANs) and finds it wins in three key areas:

Sharpness: It recovers the sharp edges of tooth roots and bone surfaces that other methods blur out.
The Hidden Tunnel: It successfully reconstructs the mandibular canal (the nerve tunnel in the jaw), a structure that previous methods completely missed or got wrong.
No Fake Teeth: It doesn't invent teeth where there are none. If a patient is missing a tooth, X-Splat leaves that space empty, whereas other methods might accidentally "hallucinate" a tooth there.

In Summary:
X-Splat is a new way to turn a flat, 2D dental X-ray into a detailed 3D model. It does this by placing millions of tiny, stretchy "balloons" along the paths of the X-ray beams, letting them grow and twist to fit the anatomy, and then making tiny, careful adjustments. The result is a 3D jaw model that is sharp, accurate, and doesn't invent fake anatomy, offering a potential low-radiation alternative to traditional 3D CT scans.

Technical Summary: X-Splat

Problem Statement

Generating a 3D dental volume from a single panoramic radiograph (PXR) presents a highly underdetermined inverse problem. Unlike Cone-Beam Computed Tomography (CBCT), which acquires hundreds of projections to reconstruct a sub-millimeter attenuation volume, a PXR integrates 3D attenuation along curved X-ray paths into a single 2D image, leaving depth-resolved anatomy unobserved.

Existing approaches to this problem face three primary limitations:

Representation Issues: Methods relying on convolutional encoder-decoders or implicit neural fields (e.g., NeRF) produce smooth representations that systematically under-resolve sharp anatomical interfaces, such as root boundaries, cortical bone surfaces, and the mandibular canal.
Input Constraints & Consistency: Some methods require additional inputs (e.g., dental arch curves) or produce volumes that are locally plausible but globally inconsistent with real dental topology.
Evaluation Deficiencies: Current evaluation relies heavily on intensity-based metrics (PSIM, SSIM) which fail to reflect whether clinically relevant structures are correctly recovered, often masking substantial structural inaccuracies.

Methodology: X-Splat

The authors propose X-Splat, the first Gaussian Splatting framework designed to generate CBCT-like 3D dental volumes from a single PXR. The method treats the known panoramic acquisition geometry as a generation scaffold, formulating the task as geometry-constrained generation rather than standard tomographic reconstruction.

Core Components

Geometry-Anchored Initialization:
- Instead of learning anchor positions from scratch, X-Splat exploits the known panoramic ray geometry.
- One learnable anisotropic 3D Gaussian primitive is initialized per ray sample (approximately 6 million anchors), densely covering the dental arch and jaw region. This concentrates representational capacity where the input is informative.
Feed-Forward Prediction & Refinement:
- Shared MLP: A single Multi-Layer Perceptron (MLP) predicts displacement, scale, rotation, and density for all Gaussian primitives in a single forward pass. The network uses global image features and per-point positional encoding.
- Constrained Movement: Gaussians are permitted to move only along their anchor rays within the axial (XY) plane to prevent cross-slice interference. They are also restricted to rotation around the superior-inferior Z-axis.
- Residual Refiner: A lightweight 3D residual U-Net refines the coarse Gaussian volume. This component adds dataset-level anatomical priors (e.g., typical jaw morphology) as small corrections without overriding the geometry resolved by the Gaussians. The refiner is kept small and initialized as a zero-correction to prevent memorization of population-level anatomy over geometric fidelity.
Training Supervision & Loss Functions:
- Beer-Lambert Reprojection: The predicted volume is rendered back to a simulated PXR and compared to the input, enforcing physical consistency.
- Multi-View DRR Consistency: To resolve depth ambiguity, the model is supervised against Digitally Reconstructed Radiographs (DRRs) from 31 azimuthal angles during training. This contracts the feasible solution set toward anatomically consistent geometries.
- Geometric & Perceptual Losses: Maximum-Intensity Projection (MIP) losses provide orthogonal structural supervision, while VGG-based perceptual losses encourage high-frequency anatomical details.
- Volumetric Intensity Loss: Mean Squared Error (MSE) is applied between the prediction and ground-truth CBCT.

Key Contributions

First Gaussian Splatting Framework for PXR-to-3D: X-Splat introduces explicit anisotropic Gaussian primitives constrained by Beer-Lambert reprojection and multi-view radiographic supervision, offering a geometry-grounded alternative to smooth implicit fields.
Recovery of Fine Structures: The method successfully recovers individual tooth roots, cortical boundaries, and the mandibular canal—structures that existing methods systematically miss or hallucinate. This is achieved through a ray-anchored representation where a shared network predicts primitive attributes based on known acquisition geometry.
Geometry-Aware Evaluation: The authors introduce segmentation-based metrics (Teeth Volume Recall, Canal Volume Recall, Hallucinated Volume) to evaluate recovery of specific maxillofacial structures, moving beyond aggregate voxel errors.

Experimental Results

The model was trained and evaluated on a subset of the ToothFairy3 dataset (60 scans with paired segmentation labels), using synthetic PXR-CBCT pairs generated via the Beer-Lambert law.

Quantitative Performance: X-Splat outperforms state-of-the-art baselines (including NeRF-based NeBLa, GANs, and Residual CNNs) across all metrics.
- Surface Accuracy: Achieved a Big Anatomy Average Surface Distance (BA-ASD) of 2.39 mm, significantly lower than the next best oracle-free method (GAN at 10.60 mm) and the implicit field baseline NeBLa (35.67 mm).
- Canal Recovery: Recovered 67.33% of the mandibular canal volume, compared to 1.14% for NeBLa and 27.62% for Residual CNN.
- Hallucination Suppression: Produced the lowest Hallucinated Volume (2.96%), indicating that multi-view DRR regularization effectively suppresses false positives.
Qualitative Performance: Visualizations show X-Splat recovers cortical and trabecular bone separation, maxillary sinuses, and the mandibular canal as continuous tubular structures. Unlike competing methods, it does not hallucinate teeth in edentulous regions.
Ablation Studies:
- Replacing the implicit field with fixed Gaussian anchors alone reduced BA-ASD by 2.9×.
- Allowing ray-aligned Gaussian movement further improved performance, acting as an implicit geometric regularizer.
- Multi-view DRR supervision ( $L_{mv}$ ) provided the single largest gain in surface accuracy, confirming its role in resolving depth ambiguity.

Significance and Claims

The paper claims that X-Splat represents a significant step toward scalable 3D dental analysis in settings where CBCT is unavailable, impractical, or unnecessary. By treating the known acquisition geometry as a scaffold, the method converts a highly underdetermined problem into a geometry-constrained volumetric inference task.

The authors emphasize that explicit Gaussian representations are particularly well-suited for dental generation, where thin cortices, root boundaries, and canal structures require localized geometric control rather than unconstrained volumetric generation.

Limitations and Future Work:
The authors note that the current study is limited to scans with full maxillofacial fields of view and paired segmentation labels, trained on synthetic data. Generalization to diverse patient populations and real-world panoramic systems (which have different spectral and intensity characteristics) remains to be established. Future work is proposed to focus on multi-site training, real-PXR adaptation, and tissue-specialized Gaussian densification.