Accelerating CEST imaging using a model-based deep neural network with synthetic training data.

Area of Science:

• Magnetic Resonance Imaging
• Biomedical Engineering
• Artificial Intelligence in Medical Imaging

Background:

• Chemical Exchange Saturation Transfer (CEST) MRI is crucial for assessing tissue physiology.
• Undersampled data acquisition in CEST MRI leads to image quality degradation.
• Developing advanced reconstruction methods is essential for efficient CEST imaging.

Purpose of the Study:

• To develop a model-based deep neural network (CEST-VN) for high-quality reconstruction of undersampled multi-coil CEST data.
• To leverage deep learning priors and multi-coil sensitivity encoding for improved CEST image reconstruction.

Main Methods:

• Unrolling the CEST image reconstruction equation into a deep neural network (CEST-VN) inspired by variational networks.
• Incorporating a k-space data-sharing block and 3D spatial-frequential convolution kernels.
• Synthesizing multi-coil CEST data using multiple-pool Bloch-McConnell simulations and training with a CEST-specific loss function.

Main Results:

• CEST-VN generated high-quality CEST source images and amide proton transfer-weighted maps in healthy and brain tumor subjects.
• The method consistently outperformed GRAPPA, blind compressed sensing, and the original variational network (VN).
• Accurate reconstruction was achieved even with increasing acceleration factors (3 to 6) without significant loss of detail or increased artifacts.

Conclusions:

• The proposed CEST-VN method enables high-quality CEST imaging from highly undersampled multi-coil data.
• Integration of deep learning priors and multi-coil sensitivity encoding models is effective.
• CEST-VN shows significant potential for improving diagnostic capabilities in neuroimaging.