Simulating Sinogram-Domain Motion and Correcting Image-Domain Artifacts Using Deep Learning in HR-pQCT Bone Imaging.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Rigid-motion artifacts in high-resolution peripheral quantitative computed tomography (HR-pQCT) impede in vivo bone microstructure assessment.
• Existing motion grading techniques lack corresponding correction methods due to absent standardized degradation models.

Purpose of the Study:

• To develop and validate a novel deep learning-based method for correcting motion artifacts in HR-pQCT images.
• To create standardized simulated motion artifacts for training supervised learning models.

Main Methods:

• Optimized a sinogram-based method to simulate motion artifacts, generating paired corrupted and ground truth HR-pQCT datasets.
• Proposed an Edge-enhanced Self-attention Wasserstein Generative Adversarial Network with Gradient Penalty (ESWGAN-GP) for motion correction.
• Integrated edge-enhancing skip connections, self-attention mechanisms, and VGG-based perceptual loss.

Main Results:

• The ESWGAN-GP model achieved high performance metrics on simulated data (SNR: 26.78, SSIM: 0.81, VIF: 0.76).
• The model demonstrated improved results on real-world data (SNR: 29.31, SSIM: 0.87, VIF: 0.81).
• The method successfully preserved trabecular edges and reconstructed fine micro-structural features.

Conclusions:

• The developed ESWGAN-GP model effectively corrects motion artifacts in HR-pQCT imaging.
• This deep learning approach represents a significant step towards mitigating motion artifacts and enhancing HR-pQCT adoption.
• While acknowledging limitations in simulating complex in vivo motion, the study provides a foundational framework for future advancements.