NeSVoR: Implicit Neural Representation for Slice-to-Volume Reconstruction in MRI

Area of Science:

• Medical Imaging
• Computational Imaging
• Machine Learning in Medical Imaging

Background:

• Reconstructing 3D MRI volumes from 2D slices is crucial for imaging moving subjects like fetuses.
• Existing methods are slow, especially for high-resolution volumes, and struggle with severe motion and artifacts.

Purpose of the Study:

• To develop a faster and more robust slice-to-volume MRI reconstruction method.
• To improve the quality and reliability of 3D MRI reconstruction in the presence of motion and artifacts.

Main Methods:

• Introduced NeSVoR, a resolution-agnostic method using implicit neural representation for continuous volume modeling.
• Incorporated a comprehensive slice acquisition model accounting for inter-slice motion, point spread function, and bias fields.
• Enabled noise variance estimation and outlier removal for enhanced reconstruction and uncertainty visualization.

Main Results:

• NeSVoR achieves state-of-the-art reconstruction quality.
• Demonstrated 2-10x acceleration in reconstruction times compared to existing methods.
• Validated performance on both simulated and in vivo data.

Conclusions:

• NeSVoR offers a significant advancement in slice-to-volume MRI reconstruction.
• The method provides high-quality, accelerated 3D volume reconstruction robust to motion and artifacts.