Multi-layer clustering-based residual sparsifying transform for low-dose CT image reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Signal Processing

Background:

• Sparsifying transform (ST) models offer computational efficiency in medical imaging.
• Deep learning models with nested structures excel at feature learning.
• Low-dose CT (LDCT) reconstruction requires advanced techniques for image quality.

Purpose of the Study:

• To propose a network-structured ST learning approach for X-ray computed tomography (CT), termed multi-layer clustering-based residual sparsifying transform (MCST) learning.
• To apply the MCST model to LDCT reconstruction by integrating it into the penalized weighted least squares (PWLS) framework.
• To enhance image reconstruction quality in LDCT by leveraging multi-layer residual maps and input clustering for accurate sparsification.

Main Methods:

• The MCST model combines a multi-layer sparse representation with clustered features, each modeled by unitary transforms.
• The model is trained unsupervisedly using a block coordinate descent (BCD) algorithm on a patch-based approach.
• A novel PWLS-MCST algorithm is developed by integrating the pre-learned MCST signal model with PWLS optimization.

Main Results:

• Experiments on phantom, numerical, and clinical LDCT datasets demonstrate the effectiveness of the MCST model.
• The learned transforms within layers capture rich features, with additional information from representation residuals.
• PWLS-MCST significantly outperforms conventional filtered back-projection (FBP) and PWLS with edge-preserving (EP) regularizers.
• The method shows superior performance compared to advanced techniques like MARS and ULTRA, particularly in edge clarity and detail preservation.

Conclusions:

• A novel multi-layer sparse signal model (MCST) with a nested network structure is introduced for CT reconstruction.
• The MCST model utilizes multi-layer residual maps and input clustering for effective image sparsification.
• The proposed PWLS-MCST framework provides clearer LDCT reconstructions, outperforming several baseline methods.
• The code for PWLS-MCST is publicly available for further research and application.