Three-Dimensional MRI Reconstruction With 3D Gaussian Representations: Tackling the Undersampling Problem.

Area of Science:

• Medical Imaging
• Computer Vision
• Signal Processing

Background:

• Three-Dimensional Gaussian representation (3DGS) shows promise in computer vision but is underexplored in magnetic resonance imaging (MRI).
• Reconstructing isotropic resolution 3D MRI from undersampled k-space data is a significant challenge in medical imaging.
• Existing MRI reconstruction techniques often require extensive training datasets or prior model training.

Purpose of the Study:

• To explore the potential of 3D Gaussian representation for reconstructing isotropic resolution 3D MRI from undersampled k-space data.
• To introduce a novel framework, 3D Gaussian MRI (3DGSMR), for explicit MR volume representation.
• To evaluate the performance of 3DGSMR compared to established 3D MRI reconstruction techniques.

Main Methods:

• Developed the 3D Gaussian MRI (3DGSMR) framework utilizing 3D Gaussian distributions for MR volume representation.
• Employed a self-supervised learning approach, eliminating the need for large training datasets or pre-trained models.
• Adapted 3DGS methodology for decomposing complex-valued MR signals.

Main Results:

• 3DGSMR effectively reconstructs voxelized 3D MR images from undersampled k-space data.
• The reconstructed image quality is comparable to well-established 3D MRI reconstruction techniques.
• The self-supervised framework enables reconstruction without prior model training.

Conclusions:

• 3D Gaussian MRI (3DGSMR) presents a novel and effective approach for 3D MRI reconstruction from undersampled data.
• The self-supervised nature of 3DGSMR offers a significant advantage by removing the dependency on extensive training data.
• This study highlights the successful adaptation of 3D Gaussian representation for magnetic resonance imaging applications.