Learnable PM diffusion coefficients and reformative coordinate attention network for low dose CT denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning methods are increasingly used for low-dose CT (LDCT) denoising.
• Aggressive denoising risks destroying critical edge and anatomical details in CT images.
• Balancing noise suppression with structure preservation remains a key challenge in LDCT denoising.

Purpose of the Study:

• To develop an advanced LDCT denoising network that effectively preserves image structures.
• To address the challenge of balancing noise reduction and edge preservation in medical imaging.

Main Methods:

• Proposed the Learnable PM diffusion coefficient and efficient attention network (PMA-Net), an encoder-decoder based deep learning model.
• Integrated a novel edge module inspired by Perona-Malik (PM) diffusion models and partial differential equations for precise edge information.
• Utilized a multiscale reformative coordinate attention module with dilated convolutions for enhanced feature extraction.
• Employed edge-enhanced multiscale perceptual loss to prevent structure loss and over-smoothing.

Main Results:

• The proposed PMA-Net demonstrated superior performance in both simulated and real-world LDCT datasets.
• Quantitative and qualitative analyses confirmed enhanced noise/artifact suppression.
• Significant improvement in preserving fine edges and anatomical structures was observed.

Conclusions:

• PMA-Net offers a novel approach to LDCT denoising by integrating partial differential equations into neural networks for edge feature extraction.
• The method achieves a superior balance between noise reduction and structural integrity.
• This work enhances the interpretability and clinical applicability of deep learning in medical imaging.