Deep learning for dense Z-spectra reconstruction from CEST images at sparse frequency offsets

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Reducing scan time in Chemical Exchange Saturation Transfer (CEST)-Magnetic Resonance Imaging (MRI) is crucial for clinical applications.
• Acquiring a sufficient number of CEST images for quantitative analysis often leads to prolonged scan times.

Purpose of the Study:

• To develop a general deep-learning framework for reconstructing dense CEST Z-spectra from sparse experimental data.
• To significantly reduce MRI scan times by minimizing the number of required CEST image acquisitions.

Main Methods:

• A sequence-to-sequence (seq2seq) deep learning framework was proposed for reconstructing dense CEST Z-spectra.
• A comprehensive training dataset was generated using simulated Z-spectra, avoiding manual annotation.
• A novel seq2seq network incorporating both short-range and long-range information processing was developed.

Main Results:

• The proposed seq2seq models accurately reconstructed dense CEST images from data acquired at only 11 sparse frequency offsets.
• Scan times for CEST-MRI were reduced by at least two-thirds.
• The novel seq2seq network demonstrated a competitive advantage in reconstruction accuracy.

Conclusions:

• Deep learning, specifically seq2seq models, offers an effective solution for reconstructing dense CEST Z-spectra from sparse data.
• The developed framework significantly reduces CEST-MRI scan time, enhancing its practical utility.
• The novel network architecture improves reconstruction capabilities, paving the way for faster and more efficient quantitative MRI.