A Dual-Encoder-Single-Decoder Based Low-Dose CT Denoising Network | JoVE Visualize

Area of Science:

Medical Imaging
Artificial Intelligence
Computer Vision

Background:

Low-dose CT (LDCT) imaging reduces radiation exposure but often suffers from noise and reduced image quality.
Generative Adversarial Networks (GANs) show promise for improving image quality in medical imaging applications.
Existing GAN architectures may not optimally process diverse feature types (shallow visual vs. deep semantic) for effective LDCT denoising.

Purpose of the Study:

To propose a novel dual-encoder-single-decoder GAN architecture for superior LDCT image quality enhancement.
To improve the generator's capability in handling both shallow visual and deep semantic features.
To enhance the discriminator's feature representation and adversarial training stability.

Main Methods:

Developed a dual-encoder-single-decoder GAN generator incorporating a pyramid non-local attention module for enhanced feature extraction.
Integrated shallow and deep feature processing modules within a secondary encoder to improve encoding capabilities.
Designed a hierarchical-split ResNet discriminator to enrich feature representation and reduce redundancy.
Employed feature fusion strategies in the generator to combine multi-level features for denoising.

Main Results:

The proposed dual-encoder-single-decoder GAN achieved superior performance in LDCT image denoising compared to traditional single-encoder GANs.
Experimental results demonstrated significant improvements in both objective image quality metrics and subjective medical diagnostic acceptability.
The enhanced feature extraction and fusion mechanisms in the generator contributed to higher quality image generation.

Conclusions:

The novel dual-encoder-single-decoder GAN effectively addresses the challenge of noise reduction in LDCT images.
The proposed architecture offers a promising approach for improving diagnostic accuracy and patient safety in CT imaging.
The method provides a valuable tool for medical image analysis and reconstruction, with code available for further research.