LAPANet is a deep learning framework for non-rigid motion estimation directly from accelerated MRI k-space data, bypassing image reconstruction. This approach enables accurate motion estimation at sub-5 millisecond temporal resolution with as few as 2 Cartesian k-space lines per frame or 3 radial spokes per frame, making it ideal for dynamic and real-time MRI applications.
✨ New Updates.
- ⏳ Pretrained weights coming soon on Hugging Face
- 🚀 Code Available for training and inference using the CMRxRecon dataset
- 🎉 Paper accepted at Medical Image Analysis: https://doi.org/10.1016/j.media.2026.104296
Highly accelerated MRI reconstructions suffer from severe undersampling artifacts and aliasing that degrade image quality and disrupt feature matching. By operating directly on acquired Fourier measurements, LAPANet estimates motion before image reconstruction, avoiding reliance on aliased images and enabling reliable motion estimation under extreme acceleration.
| Aspect | LAPANet | Image-Based Methods |
|---|---|---|
| Input | Raw k-space (accelerated) | Reconstructed images (degraded) |
| Reconstruction Needed | ❌ No | ✅ Yes |
| High Acceleration Robustness | ✅ Yes | ❌ No |
| Temporal Resolution | <5 ms | >20-50 ms |
- ✅ Direct k-space registration without image reconstruction
- ✅ Non-rigid motion estimation based on Local-All-Pass (LAP) formulation
- ✅ Self-supervised training without requiring annotated deformation fields
- ✅ Multi-coil information for complex-valued MRI data
- ✅ Multi-scale architecture capturing local and global motion patterns
- ✅ Trajectory agnostic — supports Cartesian and radial sampling
- ✅ Highly accelerated — validated at R=78 (Cartesian) and R=104 (radial)
- ✅ Real-time capable — ~30 ms inference per frame pair
- ✅ Cardiac & respiratory motion estimation validated
- Global Residual Modules — Multi-scale k-space feature extraction at full resolution
- Attention Mechanisms — Long-range spatial dependency modeling
- Motion Attention Modules — Progressive refinement across scales
- k-Space Magnitude Consistency Loss — Global structural guidance
- Efficient Architecture — 4000× speedup vs. prior LAP-based methods
Overview of the LAPANet architecture
# Clone repository
git clone https://github.com/lab-midas/LAPANet.git
cd LAPANet
# Create environment
conda env create -f environment.yml
conda activate lapanet
# Install dependencies
pip install -r requirements.txt# Download from HuggingFace (coming soon)
python scripts/download_model.py --model_name lapanet_cmrxrecon# Step-by-step inference with visualization
jupyter notebook notebooks/inference.ipynbpython hf_space/app.py| Component | Requirement | Notes |
|---|---|---|
| OS | Linux/macOS/Windows | Tested on Ubuntu 20.04+ |
| Python | 3.8–3.11 | 3.8+ recommended |
| CUDA | 11.0+ | Highly recommended for speed |
| GPU Memory | ≥8 GB | 16 GB+ for batch processing |
| RAM | ≥16 GB | 32 GB recommended |
| Disk | ≥50 GB | For datasets + checkpoints |
git clone https://github.com/lab-midas/LAPANet.git
cd LAPANet# Option A: Use provided environment (recommended)
conda env create -f environment.yml
conda activate lapanet
# Option B: Manual setup
conda create -n lapanet python=3.10
conda activate lapanet
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia# Install from requirements
pip install -r requirements.txt# not needed for cmrxrecon
git clone https://github.com/midas-tum/merlin.git
cd merlin
pip install -e .python -c "
import torch
import numpy as np
print(f'PyTorch version: {torch.__version__}')
print(f'CUDA available: {torch.cuda.is_available()}')
print(f'GPU: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"None\"}')
"# Download from CMRxRecon challenge website
# https://cmrxrecon.github.io/Home.html
# Expected directory structure:
# data/CMRxRecon/
# ├── TrainingSet/
# │ ├── AccFactor04/P001/cine_sax.mat
# │ ├── AccFactor08/P001/cine_sax.mat
# │ ├── AccFactor10/P001/cine_sax.mat
# │ └── FullSample/P001/cine_sax.mat
# └── ValidationSet/...
Edit config/train_cmrxrecon.yaml:
# Single GPU
python scripts/run_cmrxrecon.py --config config/train_cmrxrecon.yaml
# Override config parameters
python scripts/run_cmrxrecon.py \
--config configs/experiments/my_experiment.yaml \
--batch_size 64 \
--learning_rate 5e-5 \
--num_epochs 100# Using TensorBoard
tensorboard --logdir checkpoints/logs
# Using Weights & Biases (optional)
pip install wandb
# Set WANDB_API_KEY environment variableLAPANet is demonstrated in this repo using the CMRxRecon 2023 Challenge multi-coil cardiac cine dataset.
Download:
- Official website: https://cmrxrecon.github.io/Home.html
Dataset Properties:
| Property | Value |
|---|---|
| Subjects | 200 training + 100 test |
| Sequence | 2D bSSFP cine |
| Coils | 10 (multi-coil) |
| Spatial Resolution | 1.9 × 1.9 mm² |
| Temporal Phases | 25 frames |
| Slice Thickness | 8 mm |
| Acceleration Factors | 4×, 8×, 10× (Cartesian) |
| Format | MATLAB v7.3 (.mat files) |
Expected Directory Structure:
data/CMRxRecon/
├── TrainingSet/
│ ├── AccFactor04/
│ │ ├── P001/
│ │ │ ├── cine_sax.mat (undersampled k-space)
│ │ │ └── cine_lax.mat
│ │ ├── P002/...
│ │ └── ...
│ ├── AccFactor08/
│ ├── AccFactor10/
│ └── FullSample/
│ ├── P001/
│ │ ├── cine_sax.mat (fully sampled reference)
│ │ └── cine_lax.mat
│ └── ...
└── TestSet/
├── P201/ ... (similar structure)
For your own data:
- Save as HDF5/MAT/NPY with shape
(F, S, C, H, W)(complex-valued) - Or Create a custom loader
The original in-house datasets used in the paper cannot be released due to ethical restrictions.
If you use LAPANet in your research, please cite:
@article{ghoul2026learning,
title={Learning efficient non-rigid registration in k-space for accelerated Magnetic Resonance Imaging},
author={Ghoul, Aya and Hammernik, Kerstin and Lingg, Andreas and Krumm, Patrick and Rueckert, Daniel and Gatidis, Sergios and K{\"u}stner, Thomas},
journal={Medical Image Analysis},
volume={115},
pages={104296},
year={2027},
doi={10.1016/j.media.2026.104296},
publisher={Elsevier}
}- Published Article: https://doi.org/10.1016/j.media.2026.104296
- ArXiv Preprint: https://arxiv.org/abs/2410.18834
- HuggingFace Repository: https://huggingface.co/AyaGhoul/LAPANet
This project is licensed under the MIT License — see LICENSE file for details.
- Email: aya.ghoul@med.uni-tuebingen.de
- Lab Website: www.midaslab.org
- Supplementary Materials (on journal website)
Last Updated: September 2026
Current Version: 1.0.0
Code Status: 🟢 Actively Maintained