[Development of Motion Artifact Generator for Deep Learning in Brain MRI]

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Motion artifacts significantly degrade Magnetic Resonance Imaging (MRI) quality.
• Acquiring large datasets of clinical MRI with and without motion artifacts is challenging.
• Deep learning models require extensive data for effective training in artifact reduction.

Purpose of the Study:

• To develop a computer simulation method for generating realistic MRI motion artifact images.
• To create a diverse dataset of simulated artifact images for deep learning model training.
• To validate the simulation's ability to produce clinically relevant artifact characteristics.

Main Methods:

• Generated 80 distinct types of motion artifact images via computer simulation, including vertical, horizontal, diagonal, and rotational shifts.
• Transformed images into k-space data, randomly sampled phase encodings, and performed inverse Fourier transforms to create artifact images.
• Utilized a U-net deep learning model to verify the reproducibility of simulated artifact images.
• Quantitatively assessed image quality using Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR).

Main Results:

• Simulated images achieved an average SSIM of 0.95 and PSNR of 31.5.
• Clinical artifact images showed comparable results with an average SSIM of 0.96 and PSNR of 31.1.
• The simulation method demonstrated high fidelity in replicating clinical motion artifact characteristics.

Conclusions:

• The developed simulation technique efficiently generates a large volume of MRI motion artifact images rapidly.
• The simulated artifact images are equivalent in quality and characteristics to those found in clinical practice.
• This method provides a viable solution for augmenting datasets for deep learning-based MRI artifact reduction.