Motion Artifact Reduction Using U-Net Model with Three-Dimensional Simulation-Based Datasets for Brain Magnetic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Motion artifacts are a significant challenge in brain MRI, degrading image quality and potentially affecting diagnostic accuracy.
• Deep learning models, particularly U-Net architectures, show promise for image reconstruction and artifact removal.

Purpose of the Study:

• To develop and evaluate a U-Net model for the effective removal of motion artifacts from brain MR images.
• To propose a simulation method for augmenting MRI datasets to improve U-Net model training and prevent overfitting.

Main Methods:

• A U-Net model was trained using a novel simulation technique that generated motion artifacts in k-space data.
• The simulation involved 3D rotations, translations, and k-space data manipulation to create realistic motion artifacts.
• Quantitative evaluation used Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Coefficient of Correlation (CC), and Universal Image Quality Index (UQI).

Main Results:

• The U-Net model trained on the residual map-based simulated dataset demonstrated superior performance in motion artifact reduction.
• Significant improvements were observed across all quantitative metrics: RMSE (approx. 5.35×), PSNR (approx. 1.51×), CC (approx. 1.12×), and UQI (approx. 1.01×).
• The model's performance surpassed direct image comparisons, indicating effective artifact removal.

Conclusions:

• The proposed simulation-based dataset effectively trains U-Net models for robust motion artifact reduction in brain MRI.
• This method enhances the feasibility of using deep learning for improving the quality of clinical MRI scans.
• The findings suggest a pathway for developing more reliable and accurate neuroimaging techniques.