Movienet: Deep space-time-coil reconstruction network without k-space data consistency for fast motion-resolved 4D...

Area of Science:

• Medical Imaging
• Deep Learning
• Magnetic Resonance Imaging

Background:

• Dynamic MRI (4D MRI) is crucial for visualizing organ motion but is limited by long acquisition times.
• Accelerating 4D MRI acquisition and reconstruction is essential for clinical translation.

Purpose of the Study:

• To develop and validate Movienet, a deep learning approach for rapid 4D MRI reconstruction.
• To exploit space-time-coil correlations and motion preservation for accelerated golden-angle radial data acquisition.

Main Methods:

• Movienet employs a U-net architecture with residual learning in the image domain to remove aliasing.
• Motion preservation is achieved by ordering training data by motion state (expiration to inspiration).
• The method was tested on abdominal tumors using therapeutic 1.5T and diagnostic 3T MRI scanners with varying acceleration factors.

Main Results:

• Movienet achieved sub-second reconstruction times (0.69-0.75s), significantly faster than iterative methods.
• It enabled faster acquisition than XD-GRASP with comparable image quality.
• Improved suppression of streaking artifacts was observed compared to existing methods.

Conclusions:

• Movienet accelerates 4D MRI data acquisition and reconstruction, enabling sub-second imaging.
• This facilitates efficient clinical implementation for both motion-resolved 4D and motion-resistant 3D imaging.
• The deep learning approach offers a promising solution for overcoming dynamic MRI acquisition limitations.