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-magnetic resonance imaging (cine-MRI) is crucial for studying anatomical dynamics but faces reconstruction challenges due to undersampled k-space data.
• Slow MR signal acquisition limits spatial and temporal resolutions in dynamic volumetric imaging.

Approach:

• Proposed spatial and temporal implicit neural representation learning (STINR-MR) framework for accurate 3D cine-MRI reconstruction.
• Employed a joint reconstruction and deformable registration approach using spatial and temporal implicit neural networks (INRs) and principal component analysis (PCA) derived deformation vector fields (DVFs).
• Evaluated STINR-MR using simulated 4D extended cardiac-torso (XCAT) phantom data and clinical data from a healthy subject, comparing it against the MR-MOTUS method.

Key Points:

• STINR-MR reconstructs 3D cine-MR images with high temporal (<100 ms) and spatial (3 mm) resolutions.
• Demonstrated superior image quality, reduced artifacts, and enhanced tumor localization accuracy compared to MR-MOTUS.
• Achieved a mean tumor center-of-mass error of 1.0±0.4 mm in XCAT simulations, significantly better than MR-MOTUS (3.4±1.0 mm).

Conclusions:

• STINR-MR offers a lightweight, efficient, and 'one-shot' framework for 3D cine-MRI reconstruction, avoiding pre-training and generalizability issues.
• The method accurately captures irregular motion patterns due to its high-frame-rate reconstruction capability.
• STINR-MR enables precise anatomical dynamics studies and improved diagnostic accuracy in medical imaging.