CSCST-Net: a fully sparse-regularized convolutional sparse coding network for low-dose CT denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) are widely used for low-dose computed tomography (LDCT) denoising.
• However, the black-box nature of CNNs limits the interpretability of existing denoising methods.
• There is a need for interpretable and effective LDCT denoising techniques.

Purpose of the Study:

• To develop a novel, interpretable denoising model for LDCT images.
• To integrate convolutional sparse coding (CSC) with a CNN-based framework for enhanced interpretability and performance.
• To design a CNN (CSCST-Net) to solve the proposed sparse-regularized model.

Main Methods:

• Proposed a fully sparse-regularized convolutional sparse coding model (CSC-ST).
• Developed a generalized sparse transform to improve sparsity and preserve image characteristics.
• Integrated the Alternating Direction Method of Multipliers (ADMM) with gradient descent for optimization.
• Introduced adaptive convolutional dictionaries to reduce model parameters.

Main Results:

• The proposed CSCST-Net demonstrated superior performance on the Mayo Clinic dataset.
• Achieved significant improvements in noise removal and artifact suppression.
• Showcased enhanced preservation of texture details compared to state-of-the-art methods.

Conclusions:

• The CSC-ST model offers an effective and interpretable solution for LDCT denoising.
• The developed CSCST-Net shows strong advantages in practical applications.
• This approach enhances the reliability and understanding of deep learning-based medical image denoising.