Accelerated CEST imaging through deep learning quantification from reduced frequency offsets.

Area of Science:

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

Background:

• Chemical Exchange Saturation Transfer (CEST) MRI is valuable for tissue characterization.
• Acquisition time is a major limitation for CEST MRI.
• Deep learning offers potential solutions for accelerating MRI acquisition.

Purpose of the Study:

• To develop and validate a deep learning approach for constructing CEST maps from undersampled Z-spectra.
• To significantly reduce CEST MRI acquisition time while maintaining quantitative accuracy.
• To evaluate the performance of the deep learning model in simulated pathological conditions.

Main Methods:

• Fisher information gain analysis to identify optimal frequency offsets for multi-pool fitting.
• Development and training of a U-NET architecture on undersampled brain CEST data from 18 volunteers.
• Retrospective and prospective in vivo undersampling strategies were employed, reducing the number of Z-spectrum offsets.
• Simulation of glioblastoma pathology to assess network performance.

Main Results:

• Traditional multi-pool models failed to accurately quantify CEST maps from undersampled data (SSIM <0.2).
• U-NET fitting successfully generated quantitative CEST maps from undersampled Z-spectra.
• Prospective undersampling reduced scan time by 3.5 times, achieving high accuracy (MSE=4.4e-4, r=0.82, SSIM=0.84) compared to ground truth.
• The U-NET model demonstrated reliable prediction of CEST values for simulated glioblastoma.

Conclusions:

• The U-NET architecture effectively quantifies CEST maps from undersampled Z-spectra.
• This deep learning approach enables significant CEST MRI scan time reduction.
• The method shows promise for accelerated quantitative MRI in clinical settings.