Artifact estimation network for MR images: effectiveness of batch normalization and dropout layers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic resonance imaging (MRI) is crucial for medical diagnosis, but patient movement causes artifacts that degrade image quality.
• Current artifact removal methods have limitations, including extended scan times.
• Deep learning models like U-Net show potential for artifact reduction, but optimization strategies require further investigation.

Purpose of the Study:

• To develop a U-Net-based regression network for removing motion artifacts in MRI.
• To evaluate the impact of combining batch normalization (BN) and dropout layers on U-Net's performance for artifact removal.
• To compare the U-Net approach with a Transformer-based network.

Main Methods:

• A U-Net-based regression network was developed for motion artifact removal.
• Three U-Net variations were trained and tested using 1200 images (with and without artifacts).
• The study investigated the efficacy of integrating batch normalization and dropout layers into the U-Net architecture.

Main Results:

• The integration of BN and dropout layers significantly improved U-Net's accuracy in artifact removal.
• Reconstructed images showed an approximate doubling of the peak signal-to-noise ratio.
• Structural similarity index improved by approximately 10% compared to artifact-affected images.

Conclusions:

• Integrating BN and dropout layers into U-Net enhances motion artifact removal accuracy.
• The strategy shows potential for application beyond phantom images to improve MR and CT image quality.
• Optimal placement and dropout rates are crucial for maximizing the benefits of this deep learning approach.