Real-time MRI motion estimation through an unsupervised k-space-driven deformable registration network (KS-RegNet).

Area of Science:

• Medical Imaging
• Deep Learning
• Magnetic Resonance Imaging

Background:

• Real-time three-dimensional (3D) magnetic resonance (MR) imaging is hindered by slow signal acquisition, resulting in under-sampled k-space data.
• Accurate motion tracking in real-time MR imaging is crucial for various clinical applications.

Purpose of the Study:

• To propose a deep learning-based, k-space-driven deformable registration network (KS-RegNet) for real-time 3D MR imaging.
• To generate high-quality on-board MR images from under-sampled k-space data for real-time motion tracking.

Main Methods:

• KS-RegNet is an unsupervised, end-to-end network utilizing a U-Net core for deformable image registration.
• It incorporates a fully-sampled prior MR image and under-sampled k-space data of real-time MR images.
• Data fidelity loss is evaluated directly in k-space to mitigate artifacts from under-sampled images.

Main Results:

• KS-RegNet demonstrated superior and more stable performance compared to Elastix and other deep learning architectures.
• On a cardiac dataset, KS-RegNet achieved average DICE coefficients ranging from 0.884 to 0.894 and low center-of-mass errors (0.86–1.29 mm).
• The network also exhibited optimal performance on an abdominal dataset.

Conclusions:

• KS-RegNet enables real-time MRI generation with sub-second latency.
• This technology holds potential for real-time MR-guided soft tissue tracking, tumor localization, and adaptive radiotherapy.