Research map / 3D vision / Face and shape
NeBLa: Neural Beer-Lambert for 3D Reconstruction of Oral Structures from Panoramic Radiographs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeBLa, which renders simulated panoramic X-rays (SimPX) from CBCT via Beer-Lambert ray tracing, translates real PX to SimPX style with CycleGAN plus a teeth-segmentation consistency loss, and reconstructs 3D oral structures with a NeRF-inspired density MLP conditioned on a UNet image embedding.
Problem
Reconstructing 3D oral (dental and maxillofacial) structures from a single real-world 2D panoramic radiograph, without matched PX-CBCT pairs or prior information such as dental arch shape.
Approach
Generate SimPX from CBCT using PX imaging geometry (rotating ray trajectories) and the Beer-Lambert law so the synthesis is invertible; train a NeRF-style conditional density model over many patients, and bridge real PX to SimPX with unpaired CycleGAN translation regularized by a teeth segmentation consistency loss.
Key results
- On real PX to CBCT, NeBLa reaches 19.89 dB PSNR, 79.21% SSIM, 65.22% Dice and 0.30 LPIPS, beating the best baseline X2CT-GAN (18.55 dB, 77.43%, 58.85%, 0.34).
- On synthetic PX to CBCT, NeBLa reaches 21.68 dB PSNR and 74.77% Dice versus 18.55 dB and 62.55% for the best baseline (GAN/X2CT-GAN).
- Translated PX from SimPX is perceptually closer to real PX than two prior synthesis methods (AlexNet LPIPS 0.24 vs 0.33 and 0.40).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NeBLa instead of X2CT-GAN | 3D reconstruction from real PX · Real PX⇒CBCT | PSNR (dB): 18.55 -> 19.89 | Helped |
| use NeBLa instead of X2CT-GAN | 3D reconstruction from real PX · Real PX⇒CBCT | SSIM (%): 77.43 -> 79.21 | Helped |
| use NeBLa instead of X2CT-GAN | 3D reconstruction from real PX · Real PX⇒CBCT | Dice (%): 58.85 -> 65.22 | Helped |
| use NeBLa instead of X2CT-GAN | 3D reconstruction from real PX · Real PX⇒CBCT | LPIPS (VGG): 0.34 -> 0.30 | Helped |
| use NeBLa instead of the strongest baseline | 3D reconstruction from synthesized PX · Synth. PX⇒CBCT | PSNR (dB): 18.55 -> 21.68 | Helped |
| use NeBLa instead of the strongest baseline | 3D reconstruction from synthesized PX · Synth. PX⇒CBCT | SSIM (%): 78.22 -> 82.62 | Helped |
| use NeBLa instead of X2CT-GAN | 3D reconstruction from synthesized PX · Synth. PX⇒CBCT | Dice (%): 62.55 -> 74.77 | Helped |
| use NeBLa instead of X2CT-GAN | 3D reconstruction from synthesized PX · Synth. PX⇒CBCT | LPIPS (VGG): 0.32 -> 0.24 | Helped |
| add segmentation loss to the translation module (ablation) | NeBLa · Real PX⇒CBCT | Dice (%): 66.08 -> 67.00 | Helped |
| add segmentation loss to the translation module (ablation) | NeBLa · Real PX⇒CBCT | LPIPS: 0.304 -> 0.306 | Hurt |
| add projection loss to MSE-only generation loss (ablation) | NeBLa · Real PX⇒CBCT | PSNR (dB): 18.23 -> 19.61 | Helped |
| add projection loss to MSE-only generation loss (ablation) | NeBLa · Real PX⇒CBCT | LPIPS: 0.36 -> 0.37 | Hurt |
Surprising findings
- Adding the segmentation loss to the translation module gives only a small quantitative change (Dice 66.08% to 67.00%) despite clear visual improvement in teeth detail.
Limitations
- Intrinsic misalignment between PX and CBCT limits measured accuracy because the CBCT target is only approximate ground truth.
- The rotation-center trajectory of real PX equipment is unknown and approximated heuristically with a quadratic curve.
- Evaluated on a single-institution private dataset of 31 test cases.
Experimental setup
Datasets: CBCT scans from Korea University Anam Hospital (90 cases) real PX from Korea University Anam Hospital (45 unpaired cases) publicly available PX dataset for teeth segmentation
Benchmarks: Real PX to CBCT test set (31 paired cases) Synthetic PX to CBCT test set
Metrics: PSNR SSIM Dice LPIPS
Baselines: Residual CNN GAN R2N2 X2CT-GAN Oral-3D
Builds on: NeRF CycleGAN pixelNeRF X2CT-GAN
Closest papers by meaning
- Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis (NeurIPS 2025 · similarity 0.86)
- OralGPT-Plus: Learning to Use Visual Tools via Reinforcement Learning for Panoramic X-ray Analysis (CVPR 2026 · similarity 0.85)
- Segmenting Maxillofacial Structures in CBCT Volumes (CVPR 2025 · similarity 0.85)
- OralXrays-9: Towards Hospital-Scale Panoramic X-ray Anomaly Detection via Personalized Multi-Object Query-Aware Mining (CVPR 2025 · similarity 0.84)
- 3D Dental Model Segmentation with Geometrical Boundary Preserving (CVPR 2025 · similarity 0.84)
- OralGPT-Omni: A Versatile Dental Multimodal Large Language Model (CVPR 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.