An end-to-end LSTM-Attention based framework for quasi-steady-state CEST prediction

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Biomedical Engineering
• Artificial Intelligence

Background:

• Chemical Exchange Saturation Transfer (CEST) MRI requires long scan times to reach steady-state, potentially leading to underestimation of CEST effects.
• Current experimental limitations in saturation duration (Ts) and relaxation delay (Td) hinder accurate CEST measurements.

Purpose of the Study:

• To develop a deep learning model for predicting quasi-steady-state (QUASS) CEST from non-steady-state data.
• To overcome the limitations of prolonged saturation times in CEST-MRI experiments.

Main Methods:

• A multi-pool Bloch-McConnell equation was used to generate simulated Z-spectra for network training.
• A hybrid Long Short-Term Memory (LSTM)-Attention architecture was developed for QUASS CEST prediction.
• Comparative experiments evaluated LSTM-Attention against other models like MLP, RNN, GRU, and BiLSTM.

Main Results:

• The proposed LSTM-Attention model achieved superior performance in QUASS CEST prediction.
• High prediction accuracy was demonstrated with R² ≥ 0.9748, SSIM of 0.9991, PSNR of 49.6714, and MSE of 1.68×10⁻⁴.
• The model successfully predicted QUASS CEST across various frequency offsets.

Conclusions:

• The LSTM-Attention model enables high-quality QUASS CEST prediction, significantly improving upon traditional methods.
• This deep learning approach mitigates the need for extended scan durations in CEST-MRI.
• The developed model offers a time-efficient and accurate solution for CEST measurements.