Research map / 3D vision / 3D generation
DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography 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.
Learns a 3D-patch diffusion prior with position-aware positional encoding and blends slice-group scores via random adjacency and cross partitions to reconstruct 3D CT volumes from sparse-view and limited-angle measurements.
Problem
Diffusion priors for 3D CT reconstruction are hard to train on full volumes due to memory, time, and data demands, and existing 2D-slice methods rely on hand-crafted cross-slice regularization that causes z-axis inconsistency.
Approach
Train a diffusion model on joint 3D patches of k nearby slices conditioned on slice spacing via positional encoding, then at inference randomly repartition the volume each step and blend patch scores (sharing one partition per step) to form a volume-level score, with Krylov data consistency and DDIM sampling, requiring no TV regularizer.</ntml:parameter>
Key results
- On AAPM 8-view sparse-view axial reconstruction, DiffusionBlend++ reaches 35.69 dB PSNR / 0.966 SSIM versus 34.29 dB for DiffusionBlend and 33.19 dB for DDS.
- On LIDC limited-angle reconstruction (all views, axial), DiffusionBlend++ reaches 34.33 dB PSNR / 0.957 SSIM versus 30.43 dB for DiffusionBlend and 30.52 dB for DiffusionMBIR.
- DiffusionBlend++ takes about 32 minutes wall time for 8-view 3D CT reconstruction at 200 NFEs, compared with about 1400 minutes for DiffusionMBIR at 2000 NFEs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiffusionBlend++ instead of DDS | AAPM sparse-view CT reconstruction, axial, 8 views · AAPM sparse-view CT, axial, 8 views | PSNR: 33.19 -> 35.69 | Helped |
| use DiffusionBlend++ instead of DiffusionMBIR | LIDC sparse-view CT reconstruction, axial, 8 views · LIDC sparse-view CT, axial, 8 views | PSNR: 32.67 -> 34.46 | Helped |
| use DiffusionBlend++ instead of DDS | AAPM limited-angle CT reconstruction, axial · AAPM limited-angle CT, axial | PSNR: 28.20 -> 35.86 | Helped |
| use DiffusionBlend++ instead of DiffusionMBIR | LIDC limited-angle CT reconstruction, axial · LIDC limited-angle CT, axial | PSNR: 30.52 -> 34.33 | Helped |
| add adjacency-slice blending to a fixed partition (ablation) | DiffusionBlend++ · AAPM sagittal CT reconstruction | PSNR: 34.85 -> 36.02 (+1.17dB) | Helped |
| add cross-slice blending to adjacency-slice blending (ablation) | DiffusionBlend++ · AAPM sagittal CT reconstruction | PSNR: 36.02 -> 36.48 (+1.63dB reported) | Helped |
| increase jump-slice partition frequency from 2 to 32 (ablation) | DiffusionBlend++ · AAPM sparse-view CT reconstruction, axial, 8 views | PSNR: 35.69 -> 35.28 | Hurt |
| use slice blending with DiffusionBlend++ instead of DDS | AAPM sparse-view CT reconstruction, axial, 8 views, 50 NFEs · AAPM sparse-view CT, axial, 8 views, 50 NFEs | PSNR: 30.8 -> 34.5 | Helped |
Surprising findings
- DiffusionMBIR, a 2000-NFE TV-regularized baseline, is the strongest baseline on LIDC sparse-view results, beating DiffusionBlend on some sagittal and coronal cases.
- TPDM performs very poorly on the LIDC limited-angle setting (about 14 dB PSNR).
Limitations
- Experiments use only noiseless simulated measurements; robustness to noise and real measurement mismatch is untested.
- Only CT reconstruction was evaluated; other 3D inverse problems were not explored.
- Partition selection is restricted to nearby slices, so long-range inter-slice dependencies are only captured indirectly.
Experimental setup
Datasets: AAPM 2016 CT challenge LIDC-IDRI
Benchmarks: AAPM 2016 CT challenge (sparse-view and limited-angle CT) LIDC-IDRI (sparse-view and limited-angle CT)
Metrics: PSNR SSIM TV value wall time
Baselines: FBP FBP-UNet DiffusionMBIR TPDM DDS DDS 2D
Builds on: DiffusionMBIR DDS TPDM DDIM Tweedie's formula
Closest papers by meaning
- Learning Image Priors Through Patch-Based Diffusion Models for Solving Inverse Problems (NeurIPS 2024 · similarity 0.91)
- How to Use Diffusion Priors under Sparse Views? (NeurIPS 2024 · similarity 0.87)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.87)
- Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction (NeurIPS 2024 · similarity 0.87)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.87)
- Diffusion$^2$: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models (ICLR 2025 · similarity 0.86)
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.