Correction of Motion Artifacts Using a Multiscale Fully Convolutional Neural Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Motion artifacts frequently degrade Magnetic Resonance Imaging (MRI) quality in clinical settings.
• These artifacts pose a significant challenge to accurate diagnosis and interpretation of MR images.

Purpose of the Study:

• To implement and validate a novel MRI motion-artifact correction method.
• To utilize a multiscale fully convolutional neural network for artifact reduction in brain MRI.

Main Methods:

• A multiscale fully convolutional neural network was trained to identify motion artifacts in T2-weighted spin-echo brain images.
• A synthetic dataset of 93,600 images was generated using extensive data augmentation and a motion artifact simulation pipeline.
• Performance was evaluated through a blinded reader study on 28 clinical MRI cases with real patient motion.

Main Results:

• The neural network significantly improved image quality without losing morphologic information.
• A 41.84% average reduction in mean squared error was observed for synthetic test data.
• A blinded reader study showed a significant reduction in artifact scores for real-world patient data (P < .03).

Conclusions:

• Retrospective correction of motion artifacts using this multiscale fully convolutional network shows great promise.
• This method has the potential to mitigate significant motion-related problems in clinical MRI workflows.
• The approach offers a viable solution for enhancing the reliability of MRI diagnostics.