Quasi-Steady-State CEST Prediction Based on TCN-LSTM

Abstract

An important topic in chemical exchange saturation transfer (CEST)-magnetic resonance imaging (MRI) is that certain CEST effects (such as amide proton transfer effects) require sufficiently long saturation time to reach steady state, which makes CEST imaging less practical in clinical application. To address this issue, we develop a deep learning-based model to predict quasi-steady-state (QUASS) CEST from experimentally acquired CEST images with short saturation time. The study described in this paper are outlined as follows: 1) Bloch-McConnell equation is designed to obtain simulated CEST Z-spectra data, in which all possible parameters of the equation were optimized to automatically acquire large amount of training data for reflecting metabolite combinations; 2) tumor-bearing rat model was established on a 7T horizontal diameter small animal MRI scanner, allowing ground-truth generation; 3) by combining the advantages of temporal convolutional network (TCN) and long short-term memory (LSTM) in temporal modelling, a TCN-LSTM model is developed to predict QUASS CEST data. (4) To evaluate the performance of TCN-LSTM, the multilayer perceptron (MLP), recurrent neural network (RNN), LSTM, gated recurrent unit (GRU), BiLSTM and TCN are included in comparison experiment. In terms of absolute error modulus, mutual information (MI), structural similarity (SSIM) and feature similarity (FSIM), the results show that TCN-LSTM provides better prediction results than its counterparts.

Keywords

• QUASS CEST
• Bloch-McConnell equation
• deep learning