LAPANet

Learning efficient non-rigid registration in k-space for accelerated Magnetic Resonance Imaging

LAPANet is a deep learning model for fast non-rigid motion estimation in accelerated MRI. It estimates anatomically faithful motion directly from highly undersampled k-space data, without requiring additional scans or auxiliary information.

The model enables motion estimation from as little as 2 lines/frame for Cartesian MRI and 3 spokes/frame for radial MRI, with sub-5 ms temporal resolution and inference times on the order of milliseconds.

Model Details

• Model: LAPANet
• Task: Non-rigid motion estimation / image registration
• Application: Accelerated and dynamic MRI
• Input: Highly undersampled k-space data
• Output: Non-rigid motion fields
• Developed by: Aya Ghoul, Kerstin Hammernik, Andreas Lingg, Patrick Krumm, Daniel Rueckert, Sergios Gatidis, Thomas Küstner

Paper

Learning efficient non-rigid registration in k-space for accelerated Magnetic Resonance Imaging

Published in Medical Image Analysis, 2026.

📄 Paper: https://doi.org/10.1016/j.media.2026.104296

💻 Code: https://github.com/lab-midas/LAPANet

Intended Use

LAPANet is intended for research in:

• Accelerated MRI
• Dynamic MRI
• Non-rigid motion estimation
• Motion correction
• Real-time MRI applications
• MRI self-gating

The released models are intended primarily for research and should be evaluated carefully before use in clinical or other high-stakes applications.

Performance

LAPANet has demonstrated non-rigid cardiac and respiratory motion estimation from extremely undersampled acquisitions:

Acquisition	Sampling	Temporal resolution
Cartesian MRI	2 lines/frame	< 5 ms
Radial MRI	3 spokes/frame	< 5 ms

Inference is performed in the order of milliseconds, enabling potential real-time applications.

Limitations

Performance may depend on the anatomy, acquisition trajectory, sampling pattern, scanner characteristics, and data distribution. The pretrained models may not generalize to all MRI protocols or clinical scenarios.

The models have been developed and evaluated for research purposes and should not be considered clinically validated.

Citation

If you use LAPANet in your research, please cite:

@article{ghoul2026lapanet,
  title   = {Learning efficient non-rigid registration in k-space for accelerated Magnetic Resonance Imaging},
  journal = {Medical Image Analysis},
  year    = {2026},
  doi     = {10.1016/j.media.2026.104296}
}

For implementation details, pretrained model usage, and additional resources, please visit the LAPANet GitHub repository.