Enhancement based convolutional dictionary network with adaptive window for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Low-dose CT (LDCT) imaging is crucial for reducing radiation exposure.
• Conventional CNN-based denoising methods often lack interpretability and can lose image details.
• Existing approaches struggle to balance noise suppression with texture preservation in LDCT.

Purpose of the Study:

• To develop an interpretable deep learning method for effective LDCT denoising.
• To propose a novel convolutional dictionary learning model with adaptive window (CDL-AW) for enhanced noise reduction.
• To construct an enhancement-based convolutional dictionary learning network (ECDAW-Net) for improved detail retention.

Main Methods:

• Designed a CDL-AW model incorporating an adaptive window-constrained convolutional dictionary atom to minimize spectrum leakage.
• Developed ECDAW-Net by unfolding the CDL-AW model iteratively using proximal gradient descent.
• Integrated a multi-scale edge extraction module (LoG and Sobel convolutions) and a compound loss function (MSE and patch-level loss) to preserve image textures and structural information.

Main Results:

• ECDAW-Net achieved a peak signal-to-noise ratio of 33.94 and a structural similarity of 0.92 on the Mayo dataset.
• The method demonstrated superior performance in noise and artifact suppression compared to state-of-the-art techniques.
• Quantitative results indicate significant improvements in image fidelity and diagnostic quality.

Conclusions:

• The proposed interpretable ECDAW-Net effectively suppresses noise and artifacts in LDCT images.
• The method excels at preserving tissue textures and fine details, outperforming existing approaches.
• ECDAW-Net offers a promising solution for high-quality, low-dose CT imaging with enhanced interpretability.