Learning-based motion artifact removal networks for quantitative mapping

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Motion artifacts significantly degrade the quality of quantitative $$ maps derived from multi-Gradient-Recalled Echo (mGRE) MRI.
• Accurate quantitative $$ mapping is crucial for various clinical applications and research.

Purpose of the Study:

• To introduce two novel learning-based motion artifact removal networks (LEARN) for correcting motion corrupted mGRE MRI data.
• To enable the estimation of high-quality quantitative motion- and $$ -inhomogeneity-corrected $$ maps.

Main Methods:

• Two convolutional neural networks (CNNs), LEARN-IMG and LEARN-BIO, were trained to correct motion artifacts in mGRE sequences.
• LEARN-IMG corrects motion in complex mGRE images for subsequent $$ map computation.
• LEARN-BIO directly estimates motion- and $$ -inhomogeneity-corrected $$ maps from magnitude-only mGRE images using a biophysical model.

Main Results:

• Both CNNs successfully suppressed motion artifacts in synthetic and in vivo mGRE data.
• Details in the predicted quantitative $$ maps were preserved after artifact removal.
• Significant reduction in motion artifacts was observed on experimental data.

Conclusions:

• LEARN-IMG and LEARN-BIO enable high-quality quantitative $$ map estimation, corrected for motion and $$ inhomogeneity.
• LEARN-IMG processes images for subsequent analysis, while LEARN-BIO directly estimates corrected $$ maps.
• LEARN-BIO's computational speed suggests potential for broader clinical application.