Research map / 3D vision / NeRF and radiance fields
Spatiotemporal-Aware Neural Fields for Dynamic CT 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 STNF4D, a self-supervised neural field for dynamic CT that factorizes the 4D scene into four orthogonal volumes stored as hash grids and adds a spatiotemporal transformer (ST-Former) to reduce hash-conflict artifacts.
Problem
Dynamic (4D) CT reconstruction at high resolution from sparse or limited-angle projections, where existing NeRF-style plane/grid representations lack constraints and suffer from hash conflicts, causing dot-like artifacts and blur; learning-based methods need large training sets and do not generalize
Approach
Decompose the 4D scene into four orthogonal volumes mapped to compressed hash grids (concatenated rather than Hadamard-multiplied), and apply single-head self-attention across the four grid features of each sample point so the model selects features and avoids hash-conflicted ones.
Key results
- On the medical dataset, STNF4D reaches 28.60 dB PSNR / 0.9208 SSIM on average, 5.99 dB above the second-best K-plane.
- On the industrial dataset, STNF4D reaches 25.57 dB PSNR / 0.9762 SSIM, 4.26 dB above the second-best INR.
- Ablation on medical data: 4D hash grid representation raises Hex-plane baseline from 17.13 to 23.81 dB; adding ST-Former gives 28.61 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STNF4D instead of K-plane | Dynamic CT reconstruction · Medical dataset (average) | PSNR: 5.99 dB | Helped |
| Use STNF4D instead of TT-U-Net | Dynamic CT reconstruction · Medical dataset (average) | SSIM: 9.95% | Helped |
| Use STNF4D instead of INR | Dynamic CT reconstruction · Industrial dataset (average) | PSNR: 4.26 dB | Helped |
| Use STNF4D instead of K-plane | Dynamic CT reconstruction · Industrial dataset (average) | SSIM: 3.23% | Helped |
| Add 4D hash grid representation instead of six-plane decomposition (ablation) | STNF4D without ST-Former · Medical dataset | PSNR: 17.13 -> 23.81; 6.68 dB | Helped |
| Add ST-Former (ablation) | STNF4D with 4D hash grid representation · Medical dataset | PSNR: 23.81 -> 28.61; 4.11 dB | Helped |
Surprising findings
- Simply increasing the hash table size yields only limited PSNR gains because grid features lack constraints.
Limitations
- ST-Former uses single-head attention without multi-head or complex attention, chosen for efficiency with many sample points.
- Per-scene optimization; learning-based baselines are not directly comparable in training setup.
- Medical and industrial results are reported on synthesized projections from a limited set of cases.
Experimental setup
Datasets: XCAT phantom 4D-Lung Cancer Imaging Archive (TCIA) aluminum product 4D-CT dataset (Reed et al. 2021)
Benchmarks: XCAT phantom and TCIA medical 4D-CT aluminum industrial 4D-CT
Metrics: PSNR SSIM
Baselines: Gate-FDK PICCS INR K-plane Hex-plane CycN-Net TT-U-Net
Builds on: Instant-NGP K-planes Hex-plane NeRF
Closest papers by meaning
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.85)
- ReMatching Dynamic Reconstruction Flow (ICLR 2025 · similarity 0.85)
- ResFields: Residual Neural Fields for Spatiotemporal Signals (ICLR 2024 · similarity 0.85)
- Towards Prospective Medical Image Reconstruction via Knowledge-Informed Dynamic Optimal Transport (NeurIPS 2025 · similarity 0.84)
- Grid4D: 4D Decomposed Hash Encoding for High-Fidelity Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.84)
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering (CVPR 2024 · similarity 0.84)
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.