A dictionary-based graph-cut algorithm for MRI reconstruction.

Area of Science:

• Medical Imaging
• Signal Processing
• Computer Vision

Background:

• Compressive sensing (CS) image reconstruction relies on random undersampling to create aliasing artifacts.
• Sparsity-enforcing priors, particularly L_p norms (0 ≤ p ≤ 1), are crucial for denoising in CS.
• Existing CS methods often use convex relaxations like L₁ norm, potentially limiting reconstruction power.

Purpose of the Study:

• To develop an efficient discrete optimization formulation for CS image reconstruction.
• To incorporate non-convex, edge-preserving priors beyond standard L_p norms.
• To address the challenges of highly non-convex minimization problems in advanced CS.

Main Methods:

• The study combines edge-preserving priors with random undersampling techniques.
• A discrete optimization approach utilizing graph cuts is employed to solve the reconstruction problem.
• Graph cuts are applied iteratively within a dictionary tailored for brain MRI data.

Main Results:

• The proposed algorithm demonstrates superior performance in reconstructing in vivo data.
• Experimental results validate the effectiveness of the novel approach.
• The method outperforms traditional regularized SENSE and standard CS reconstruction.

Conclusions:

• The developed discrete optimization method enhances compressive sensing image reconstruction.
• The use of non-convex, edge-preserving priors and graph cuts leads to improved results.
• This approach offers a significant advancement for in vivo MRI data reconstruction.