Research map / 3D vision / Depth and stereo
Fully Convolutional Slice-to-Volume Reconstruction for Single-Stack MRI
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 a fully convolutional Splat–Slice U-Net that predicts slice motion from a single MRI slice stack, using splatting/slicing and a rigid motion-compensating loss, with a U-Net interpolator to fill holes in the reconstruction.
Problem
Slice-to-volume reconstruction (SVR) of motion-corrupted 2D MRI slice stacks normally needs multiple orthogonal stacks, lengthening scans and limiting use in time-sensitive settings such as fetal fMRI.
Approach
Formulate SVR as supervised single-stack motion estimation: a 3D FCN (Splat–Slice U-Net) interprets classical SVR's slice/splat operations to predict per-slice dense motion fields, with the 3D volume obtained by splatting slices using the predicted motion, plus a rigid motion-compensated MSE loss and a separate U-Net interpolator.
Key results
- Single-stack motion prediction error of 1.81 mm APE/EPE versus 3.27 mm for SVoRT (v2) and 8.08 mm for SVRnet, a 77.6% and 44.6% reduction relative to SVRnet and SVoRT respectively.
- Volume PSNR of 23.43 dB versus 18.49 dB for SVoRT (v2) and 16.03 dB for SVRnet on fetal validation.
- Inference takes 0.224 s per stack, versus 0.011 s for SVRnet and 0.142 s for SVoRT (v2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed single-stack FCN instead of SVoRT (v2) | Fetal brain SVR · Fetal brain validation, 1 stack | Motion EPE (mm): 3.27±0.71 -> 1.81±0.40 | Helped |
| Use the proposed single-stack FCN instead of SVoRT (v2) | Fetal brain SVR · Fetal brain validation, 1 stack | Slice PSNR (dB): 20.77±1.28 -> 23.69±1.39 | Helped |
| Use the proposed single-stack FCN instead of SVoRT (v2) | Fetal brain SVR · Fetal brain validation, 1 stack | Volume PSNR (dB): 18.49±1.63 -> 23.43±1.40 | Helped |
| Use the proposed single-stack FCN instead of SVoRT (v2) | Fetal brain SVR · Fetal brain validation, 1 stack | Time (sec): 0.142s -> 0.224s | Hurt |
| Compensate for global rigid motion in the training loss (ablation) | Splat–Slice U-Net motion prediction · Adult brain validation | Motion prediction accuracy | Helped |
| Compensate for translations only instead of global rigid motion (ablation) | Splat–Slice U-Net motion prediction · Adult brain validation | Motion prediction accuracy | Hurt |
| Use ordinary MSE loss without motion compensation (ablation) | Splat–Slice U-Net motion prediction · Adult brain validation | Motion prediction accuracy | Hurt |
| Train separate models for different slice directions (ablation) | Splat–Slice U-Net · Adult brain validation, sagittal versus axial stacks | Motion MSE (voxels): 0.89 (MSE) -> 1.33 (MSE) | Mixed |
Surprising findings
- The SVRnet baseline is slower but its runtime (0.011 s) is faster than the proposed method (0.224 s); the proposed method trades a little speed for accuracy.
Limitations
- Fetal SVR network trained on automatically segmented brain slices, not raw unsegmented stacks.
- Only rigid motion is handled; deformable motion (e.g., fetal torso, placenta) is future work.
- Reconstruction can contain holes requiring a separate interpolation network.
Experimental setup
Datasets: ABIDE ADHD200 COBRE GSP MCIC OASIS PPMI UKB Buckner39 CRL Fetal Atlases FeTA MIAL
Benchmarks: FeTA (validation, synthetic slicing) MIAL real fetal acquisitions adult brain synthetic slicing
Metrics: motion MSE end-point error (EPE) slice PSNR volume PSNR APE runtime
Baselines: SVRnet SVoRT (v2) PlaneInVol
Builds on: U-Net Rousseau et al. slice-to-volume reconstruction monocular disparity estimation SVoRT SVRnet
Closest papers by meaning
- SVRMamba: Slice-to-Volume Reconstruction from Multiple MRI Stacks with Slice Sequence Guided Mamba (AAAI 2025 · similarity 0.91)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.86)
- LiFT: Lifted Inter-slice Feature Trajectories for 3D Image Generation from 2D Generators (NeurIPS 2026 · similarity 0.85)
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.85)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
- Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI (CVPR 2024 · similarity 0.85)
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.