3D cine-magnetic resonance imaging using spatial and temporal implicit neural representation learning (STINR-MR)

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Anatomy

Background:

• 3D cine-MRI is crucial for studying anatomical dynamics but faces challenges due to under-sampled k-space data.
• Reconstructing high-resolution dynamic volumetric images requires advanced techniques to overcome acquisition speed limitations.

Purpose of the Study:

• To develop and evaluate a novel machine learning framework, STINR-MR, for accurate 3D cine-MRI reconstruction from highly under-sampled data.
• To improve spatial and temporal resolution in dynamic MRI while reducing artifacts.

Main Methods:

• STINR-MR employs a joint reconstruction and deformable registration approach using spatial and temporal implicit neural representations (INRs).
• It reconstructs a reference-frame 3D MR image and a motion model, utilizing basis deformation vector fields (DVFs) derived from principal component analysis.
• A temporal INR encodes time points to generate time-resolved motion fields for cine-frame-specific dynamics.

Main Results:

• STINR-MR successfully reconstructed 3D cine-MR images with high temporal (<100 ms) and spatial (3 mm) resolutions.
• The method demonstrated superior image quality and reduced artifacts compared to MR-MOTUS and TEMPEST.
• It achieved significantly improved tumor localization accuracy, with a mean error of 0.9 ± 0.4 mm in XCAT phantom studies.

Conclusions:

• STINR-MR offers a lightweight, efficient, and 'one-shot' framework for accurate 3D cine-MRI reconstruction.
• The method avoids generalizability issues common in deep learning by not requiring external pre-training data.
• STINR-MR enables accurate capture of irregular motion patterns, advancing dynamic volumetric imaging.