Research map / 3D vision / Depth and stereo
SVRMamba: Slice-to-Volume Reconstruction from Multiple MRI Stacks with Slice Sequence Guided Mamba
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 SVRMamba, a Mamba-based slice-to-volume reconstruction method that uses odd-even slice-sequence scanning with sequence embeddings to predict slice motion, followed by a 3D U-Net interpolation network, achieving ~60x faster reconstruction than SOTA.
Problem
Fetal MRI slice-to-volume reconstruction from multiple 2D slice stacks suffers from slice misalignment and noise; existing coarse-to-fine SVR methods are slow and fail to model spatial relationships between slices.
Approach
Reformulates Mamba's unidirectional scan into a slice-sequence-guided odd-even directional scan with learnable slice/stack sequence embedding tokens to regress slice motion in one pass, and replaces iterative volume refinement with a CNN (3D U-Net) interpolation network that fills holes and suppresses noise.
Key results
- SVRMamba achieves PSNR 23.84 / SSIM 0.83 on SBA, 23.46 / 0.83 on FeTA, and 23.61 / 0.84 on FBA, outperforming all SOTA baselines.
- SVRMamba reconstructs in about 3.0 s versus 183.9 s for NeSVoR, a sixtyfold speedup with comparable accuracy.
- Ablations show removing slice sequence scan (SSC), sequence embedding (SSE), or the interpolation network (IN) degrades PSNR across all three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use slice sequence-guided odd-even scanning (SSC) (ablation) | SVRMamba · SBA | PSNR: 23.41 -> 23.84 | Helped |
| add slice sequence embedding (SSE) (ablation) | SVRMamba · SBA | PSNR: 22.68 -> 23.84 | Helped |
| use the CNN interpolation network (IN) instead of PSF-based interpolation (ablation) | SVRMamba · SBA | PSNR: 22.46 -> 23.84 | Helped |
| use slice sequence-guided odd-even scanning (SSC) (ablation) | SVRMamba · FeTA | PSNR: 22.59 -> 23.46 | Helped |
| add slice sequence embedding (SSE) (ablation) | SVRMamba · FeTA | PSNR: 20.64 -> 23.46 | Helped |
| use the CNN interpolation network (IN) instead of PSF-based interpolation (ablation) | SVRMamba · FeTA | PSNR: 21.73 -> 23.46 | Helped |
| use slice sequence-guided odd-even scanning (SSC) (ablation) | SVRMamba · FBA | PSNR: 22.95 -> 23.61 | Helped |
| add slice sequence embedding (SSE) (ablation) | SVRMamba · FBA | PSNR: 21.72 -> 23.61 | Helped |
| use the CNN interpolation network (IN) instead of PSF-based interpolation (ablation) | SVRMamba · FBA | PSNR: 22.46 -> 23.61 | Helped |
Limitations
- Trained only on simulated motion applied to registered ground-truth volumes; clinical evaluation is qualitative only.
- Interpolation network choice is not optimized; the authors state searching for the best interpolation model is not the focus.
Experimental setup
Datasets: FeTA FBA SBA
Benchmarks: FeTA FBA SBA clinical T2-weighted fetal MRI datasets
Metrics: PSNR SSIM MAE translation MAE rotation reconstruction time
Baselines: NiftyMIC SVRTK NeSVoR SVoRT+NeSVoR
Builds on: Mamba Vision Mamba VMamba NiftyMIC NeSVoR
Closest papers by meaning
- Fully Convolutional Slice-to-Volume Reconstruction for Single-Stack MRI (CVPR 2024 · similarity 0.91)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.83)
- LiFT: Lifted Inter-slice Feature Trajectories for 3D Image Generation from 2D Generators (NeurIPS 2026 · similarity 0.83)
- CycleINR: Cycle Implicit Neural Representation for Arbitrary-Scale Volumetric Super-Resolution of Medical Data (CVPR 2024 · similarity 0.82)
- LMO: Linear Mamba Operator for MRI Reconstruction (CVPR 2025 · similarity 0.82)
- S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation (AAAI 2025 · similarity 0.81)
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.