2D probabilistic undersampling pattern optimization for MR image reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• 3D Magnetic Resonance Imaging (MRI) faces a trade-off between image quality and scan time.
• Reconstructing high-quality MRI from undersampled k-space is a key challenge.
• Existing methods often optimize undersampling patterns and reconstruction separately, limiting performance.

Purpose of the Study:

• To develop a joint optimization model for simultaneously optimizing k-space undersampling patterns and MRI reconstruction.
• To improve the quality and efficiency of 3D MRI acquisition.
• To explore the relationship between undersampling probability distribution and sampling rate.

Main Methods:

• A novel end-to-end trained joint optimization model was proposed.
• A 2D probabilistic undersampling layer was designed for differentiable optimization of undersampling patterns.
• A 2D inverse Fourier transform layer connected the Fourier and image domains.
• The model was tested on 3D T1-weighted brain MR images from public and local datasets.

Main Results:

• The proposed 2D probabilistic undersampling pattern significantly outperformed state-of-the-art undersampling strategies.
• Both qualitative and quantitative comparisons demonstrated superior performance.
• An optimized relationship between undersampling probability distribution and sampling rate was identified.

Conclusions:

• Joint optimization of undersampling patterns and reconstruction models offers significant advantages for 3D MRI.
• The proposed method shows strong performance and generalization capabilities.
• This approach advances the field of accelerated MRI acquisition.