Structure-preserving low-dose computed tomography image denoising using a deep residual adaptive global context...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) reduces radiation but degrades image quality.
• Conventional deep convolutional neural networks (CNNs) struggle with nonlocal structures and regional statistics in CT images.
• This limitation hinders effective denoising for LDCT scans.

Purpose of the Study:

• To propose an adaptive global context (AGC) modeling scheme for CT images.
• To develop an AGC-based long-short residual encoder-decoder (AGC-LSRED) network for LDCT image denoising.
• To enhance noise artifact suppression in LDCT images.

Main Methods:

• Introduced an adaptive global context (AGC) modeling scheme to capture nonlocal correlations and regional statistics.
• Developed an AGC-based long-short residual encoder-decoder (AGC-LSRED) network utilizing residual AGC attention blocks (RAGCBs).
• Employed long and short skip connections within the AGC-LSRED network to facilitate deep network training.
• Utilized a compound loss function combining L1 loss, adversarial loss, and self-supervised multi-scale perceptual loss for training.

Main Results:

• The proposed method achieved superior noise suppression (RMSE = 9.02, PSNR = 33.17) and fine structure preservation (SSIM = 0.925) in simulation experiments.
• Real-data experiments showed the method outperformed others in subjective assessments by radiologists (average score = 4.34).
• Demonstrated effectiveness in noise reduction and preservation of anatomical details.

Conclusions:

• The AGC modeling scheme effectively characterizes structural information in CT images.
• Residual AGC-attention blocks with skip connections ease network training for denoising.
• The AGC-LSRED method successfully suppresses noise and preserves fine anatomical structures in LDCT images.