Low dose CT reconstruction via L1 norm dictionary learning using alternating minimization algorithm and balancing...

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Excessive radiation exposure from computed tomography (CT) scans poses a significant clinical concern due to increased cancer risk.
• Statistical iterative reconstruction (SIR) with l0-norm dictionary learning has been used for low-dose CT image reconstruction, but its l0-norm penalty lacks guaranteed global convergence.

Purpose of the Study:

• To address the convergence limitations of l0-norm regularization in low-dose CT.
• To develop a novel SIR framework utilizing l1-norm dictionary learning for improved CT image reconstruction.

Main Methods:

• Introduced an l1-norm dictionary learning penalty into the SIR framework.
• Developed an alternating minimization algorithm to solve the objective function, separating it into sparse coding and image updating subproblems.
• Employed an efficient model function approach for regularization parameter selection during image updating.

Main Results:

• The proposed l1-norm based SIR algorithm demonstrated comparable performance to the l0-norm approach.
• Quantitative metrics (RMSE, SSIM) and visual assessments confirmed the algorithm's effectiveness.
• Outperformed conventional filtered backprojection (FBP) and total variation (TV) minimization methods.

Conclusions:

• The l1-norm dictionary learning penalty offers a viable alternative for low-dose CT image reconstruction within the SIR framework.
• The developed alternating minimization algorithm provides robust performance, comparable to existing methods while addressing convergence issues.