Low-dose CT reconstruction via L1 dictionary learning regularization using iteratively reweighted least-squares.

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Compressed sensing enables high-quality computed tomography (CT) reconstruction from sparse data, reducing radiation dose.
• Dictionary learning (DL) algorithms have been developed for sparse CT reconstruction.
• Existing DL methods using L2-norm regularization show limitations in reconstruction quality at lower sampling rates.

Purpose of the Study:

• To improve dictionary learning-based CT reconstruction for further radiation dose reduction.
• To address the deteriorating reconstruction quality of existing DL methods at low sampling rates.

Main Methods:

• Replaced the L2-norm regularization term with an L1-norm term in the DL algorithm.
• Developed the L1-DL method to alleviate over-smoothing and preserve image details.
• Solved the L1-minimization problem using a weighting strategy and iteratively reweighted least squares (IRLS).

Main Results:

• The proposed L1-DL algorithm demonstrated higher accuracy compared to existing DL (ADSIR) and other compressed sensing methods.
• Improved performance was observed particularly at reduced sampling rates and increased noise levels.
• Numerical simulations confirmed the superiority of the L1-DL approach.

Conclusions:

• The L1-DL algorithm effectively utilizes image sparsity prior information.
• Replacing L2-norm with L1-norm regularization and employing IRLS strategy enhances image reconstruction accuracy.
• The L1-DL method provides a more exact reconstruction for low-dose CT applications.