Gradient Guided Co-Retention Feature Pyramid Network for LDCT Image Denoising.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but introduces image noise and artifacts.
• Conventional Feature Pyramid Networks (FPNs) struggle to balance spatial resolution and semantic value in feature extraction.
• Existing methods often generalize details in upper layers, diminishing fine image features.

Purpose of the Study:

• To propose a novel network, the Gradient Guided Co-Retention Feature Pyramid Network (G2CR-FPN), for enhancing LDCT images.
• To address the trade-off between spatial resolution and semantic information in feature maps derived from LDCT scans.
• To improve the quality of CT images obtained with reduced radiation doses.

Main Methods:

• The G2CR-FPN employs a three-path structure: bottom-up, lateral, and top-down.
• The bottom-up path generates hierarchical feature maps using FPN principles.
• The lateral path integrates directional gradient approximations for edge enhancement, while the top-down path uses a co-retention block guided by these gradients to preserve semantic value.

Main Results:

• Experimental results on clinical CT images demonstrated the effectiveness of the G2CR-FPN.
• The proposed network successfully addressed the challenge of maintaining both spatial detail and semantic richness in LDCT feature maps.
• The G2CR-FPN showed promising performance in improving the quality of low-dose CT images.

Conclusions:

• The G2CR-FPN offers a significant advancement in processing LDCT images by effectively managing the spatial-semantic trade-off.
• This novel approach holds potential for enhancing diagnostic accuracy in medical imaging applications utilizing low-dose CT scans.
• The developed G2CR-FPN framework provides a robust solution for noise and artifact reduction in low-dose CT imaging.