Research map / 3D vision / Depth and stereo
MotionTTT: 2D Test-Time-Training Motion Estimation for 3D Motion Corrected 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 MotionTTT, a test-time-training method that estimates 3D rigid motion parameters by optimizing a data-consistency loss through a 2D motion-free-trained U-net, then reconstructs motion-corrected MRI volumes.
Problem
Patient motion during long 3D MRI acquisitions causes severe artifacts; existing deep learning methods mostly handle 2D in-plane motion or rely on simulated motion during training, and classical alternating optimization is slow and inaccurate under severe motion.
Approach
A U-net trained only on motion-free 2D slices has low data-consistency loss when the motion parameters correctly undo the corruption; so optimizing the motion parameters at test time through the frozen network (slice-wise over 3D volumes, with a three-phase schedule and DC-loss thresholding) estimates 3D rigid motion without motion-corrupted training data.
Key results
- MotionTTT with estimated motion matches known-motion PSNR across motion severities for L1 reconstruction, while alternating optimization degrades under severe motion.
- MotionTTT is about 6x faster than alternating optimization on the simulated inter-shot task.
- On real prospectively acquired motion data, MotionTTT yields significantly reduced motion artifacts versus no correction (visual assessment).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the combined image-domain and k-space pre-training loss instead of image-domain-only loss (ablation) | U-Net pre-training and MotionTTT · Motion-free reconstruction | PSNR: 36.64 -> 36.73 | Helped |
| Use the combined image-domain and k-space pre-training loss instead of k-space-only loss (ablation) | U-Net pre-training and MotionTTT · Motion-free reconstruction | PSNR: 36.59 -> 36.73 | Helped |
| Use the combined image-domain and k-space pre-training loss instead of k-space-only loss (ablation) | MotionTTT-Th-L1 · Inter-shot motion, severity 9 | PSNR: 35.23261 -> 35.23587 | Helped |
| Estimate intra-shot motion in Phase 3 rather than using Phase 1 shot-level estimates (ablation) | MotionTTT-L1 · Simulated intra-shot motion | PSNR | Helped |
| Apply DC-loss thresholding to discard poorly estimated motion states (ablation) | MotionTTT-L1 · Simulated inter-shot motion, strong severity | PSNR | Helped |
| Use L1-based reconstruction instead of U-Net reconstruction (ablation) | MotionTTT reconstruction · Simulated inter-shot motion, larger motion severity | PSNR | Helped |
| Use U-Net reconstruction instead of L1-based reconstruction (ablation) | MotionTTT reconstruction · Simulated inter-shot motion, no motion | PSNR | Helped |
| Estimate motion with MotionTTT instead of alternating optimization | 3D MRI motion estimation and reconstruction · Simulated inter-shot motion, severe motion | PSNR | Helped |
| Estimate motion with MotionTTT instead of alternating optimization | 3D MRI motion estimation · Simulated inter-shot motion | runtime: about 6x faster | Helped |
| Apply MotionTTT motion correction instead of no motion correction | 3D MRI reconstruction · Prospectively acquired real motion-corrupted data | visual reconstruction quality | Helped |
| Set the intra-shot split count to 10 (ablation) | MotionTTT intra-shot motion estimation · Simulated intra-shot motion | motion estimation performance | Mixed |
| Acquire k-space lines in random order within each shot (ablation) | MotionTTT intra-shot motion estimation · Simulated intra-shot motion | motion estimation performance | Helped |
Surprising findings
- L1-based reconstruction outperforms the U-net reconstruction under strong motion, attributed to distribution shift in effective undersampling masks.
Limitations
- Reconstruction module uses L1-minimization; deep reconstruction not yet explored.
- Network trained on fully-sampled data, which is scarce for 3D high resolution.
- Real-data evaluation is limited to one subject with four scans.
Experimental setup
Datasets: Calgary-Campinas Brain MRI Dataset
Benchmarks: Calgary-Campinas Brain MRI (simulated motion) prospectively acquired real motion data
Metrics: PSNR DC loss visual quality
Baselines: AltOpt-L1 (alternating optimization) L1-minimization without motion estimation KnownMotion-L1 KnownMotion-U-net-DCLayer
Builds on: U-net alternating optimization motion correction ESPIRiT Adam NUFFT
Closest papers by meaning
- Moner: Motion Correction in Undersampled Radial MRI with Unsupervised Neural Representation (ICLR 2025 · similarity 0.85)
- D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction (NeurIPS 2025 · similarity 0.83)
- Spatiotemporal-Aware Neural Fields for Dynamic CT Reconstruction (AAAI 2025 · similarity 0.83)
- Fully Convolutional Slice-to-Volume Reconstruction for Single-Stack MRI (CVPR 2024 · similarity 0.83)
- Towards Prospective Medical Image Reconstruction via Knowledge-Informed Dynamic Optimal Transport (NeurIPS 2025 · similarity 0.83)
- Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · 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.