Deep learning-assisted model-based off-resonance correction for non-Cartesian SWI.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Image Reconstruction
• Deep Learning in Medical Imaging

Background:

• Patient-induced magnetic field inhomogeneities cause artifacts (off-resonance) in MRI, particularly in long readout sequences like SWI.
• Conventional correction methods are too slow for high-resolution, accelerated 3D non-Cartesian multi-coil acquisitions in clinical settings.

Purpose of the Study:

• To develop and evaluate a hybrid pipeline combining model-based and neural network approaches for accelerated off-resonance correction in MRI.
• To reduce the computational time of off-resonance correction in demanding MRI acquisitions.

Main Methods:

• Hybrid pipelines using UNets were trained on accelerated 3D SWI SPARKLING acquisitions.
• Off-resonance correction was performed using a combination of model-based methods and deep learning.
• Performance was compared against model-only and network-only pipelines using slow, model-based corrections as ground truth.

Main Results:

• The proposed hybrid pipelines achieved reconstruction speeds two to three times faster than baseline methods.
• Neural networks acted as a pre-conditioner and provided inter-iteration memory, enhancing model design flexibility.
• Synergies between acceleration factors and model/network components were observed.

Conclusions:

• A combination of model-based and network-based off-resonance correction effectively accelerates conventional methods.
• The hybrid approach shows promise for clinical application by significantly reducing reconstruction times.
• Further exploration of model/network synergies could lead to future advancements in accelerated MRI.