Research map / 3D vision / 3D generation
C^2RV: Cross-Regional and Cross-View Learning for Sparse-View CBCT Reconstruction
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 C2RV, a DIF-Net-based implicit-field framework for sparse-view CBCT reconstruction that adds multi-scale 3D volumetric features and scale-view cross-attention to aggregate voxel- and pixel-aligned features.
Problem
Sparse-view cone-beam CT reconstruction from few 2D projections performs poorly on complicated anatomies, because prior implicit-neural-representation methods use only local pixel-aligned features and treat all views equally.
Approach
Back-project multi-scale 2D encoder features into explicit 3D volumes (MS-3DV) to enable cross-regional learning, and aggregate multi-scale voxel-aligned and multi-view pixel-aligned features with a scale-view cross-attention module that adaptively weights views and scales.
Key results
- On LUNA16 chest CT with 6 views, C2RV reaches 29.23 dB PSNR / 92.78 SSIM versus 25.55 / 84.40 for DIF-Net.
- On knee CBCT with 10 views, C2RV reaches 31.55 dB PSNR / 95.01 SSIM versus 29.33 / 92.06 for DIF-Net.
- Left-lung Dice on 6-view chest reconstructions improves from 94.45% (DIF-Net) to 96.72% (C2RV).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DIF-Net with C2RV | DIF-Net · LUNA16 chest CT, 6-View | PSNR (dB): 25.55 -> 29.23 | Helped |
| replace DIF-Net with C2RV | DIF-Net · LUNA16 chest CT, 8-View | PSNR (dB): 26.09 -> 29.95 | Helped |
| replace DIF-Net with C2RV | DIF-Net · LUNA16 chest CT, 10-View | PSNR (dB): 26.67 -> 30.70 | Helped |
| replace DIF-Net with C2RV | DIF-Net · Lin et al. knee CBCT, 6-View | PSNR (dB): 27.12 -> 29.73 | Helped |
| replace DIF-Net with C2RV | DIF-Net · Lin et al. knee CBCT, 8-View | PSNR (dB): 28.31 -> 30.68 | Helped |
| replace DIF-Net with C2RV | DIF-Net · Lin et al. knee CBCT, 10-View | PSNR (dB): 29.33 -> 31.55 | Helped |
| add MS-3DV (ablation) | DIF-Net · LUNA16 chest CT, 6-View | PSNR (dB): 25.55 -> 26.62 | Helped |
| replace MLP aggregation with SVC-Att (ablation) | DIF-Net · LUNA16 chest CT, 6-View | PSNR (dB): 25.55 -> 27.84 | Helped |
| add MS-3DV and SVC-Att (ablation) | DIF-Net · LUNA16 chest CT, 6-View | PSNR (dB): 25.55 -> 29.23 | Helped |
| use 1 scale instead of 3 (ablation) | C2RV · LUNA16 chest CT, 6-View | PSNR (dB): 29.23 -> 28.98 | Hurt |
| use 2 scales instead of 3 (ablation) | C2RV · LUNA16 chest CT, 6-View | PSNR (dB): 29.23 -> 29.09 | Hurt |
| use decoder output as F1 instead of encoder output (ablation) | C2RV · LUNA16 chest CT, 6-View | PSNR (dB): 29.23 -> 28.57 | Hurt |
Surprising findings
- DIF-Net performs well on knee CBCT but degrades substantially on the more complex chest anatomy, motivating the multi-scale design.
Limitations
- Performance drops on unseen anatomies (e.g., chest to head), since the model learns dataset-specific priors.
- Sensitive to noisy scanning parameters: 3 mm DSO noise lowers PSNR by 1.38 dB.
- Evaluated only on chest and knee datasets, with dental results only in the supplement.
Experimental setup
Datasets: LUNA16 Knee CBCT dataset (Lin et al. 2023)
Benchmarks: LUNA16 Knee CBCT dataset
Metrics: PSNR SSIM Dice ASD
Baselines: FDK SART NAF NeRP FBPConvNet FreeSeed BBDM PixelNeRF DIF-Net
Builds on: DIF-Net PixelNeRF U-Net multi-head cross-attention
Closest papers by meaning
- Unsupervised Self-Prior Embedding Neural Representation for Iterative Sparse-View CT Reconstruction (AAAI 2025 · similarity 0.86)
- Segmenting Maxillofacial Structures in CBCT Volumes (CVPR 2025 · similarity 0.84)
- QN-Mixer: A Quasi-Newton MLP-Mixer Model for Sparse-View CT Reconstruction (CVPR 2024 · similarity 0.84)
- DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction (NeurIPS 2024 · similarity 0.83)
- R$^2$-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction (NeurIPS 2024 · similarity 0.83)
- Sparse Bayesian Deep Learning for Cross Domain Medical Image Reconstruction (AAAI 2024 · 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.