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

Wei Yang, Jisheng Zou, Xuan Zhang, Yaowen Chen, Hanjing Tang, Gang Xiao, Xiaolei ZhangFrontiers in Neuroscience 2024 · DOI

TL;DR

Researchers have developed a hybrid deep learning model, LSTM-Attention, that predicts high-quality "quasi-steady-state" (QUASS) MRI images from short, non-steady-state scans. By training on massive datasets generated via physical simulations, the framework allows for accurate molecular imaging while significantly reducing the time patients must spend inside the scanner.

What Problem Does This Paper Address?

Chemical Exchange Saturation Transfer (CEST) is a powerful MRI technique that allows clinicians to map specific molecules (like proteins or metabolites) in the body. However, CEST has a major "speed limit." To get accurate, reproducible measurements, the scanner needs to reach a "steady state," which requires a long saturation duration ($T_s$) and relaxation delay ($T_d$).

In a clinical setting, these long scan times are impractical—they lead to patient discomfort and motion artifacts. If clinicians cut the scan short to save time, the resulting measurements are underestimated and vary between different MRI machines. This paper addresses the challenge of "recovering" the accurate steady-state information from much faster, non-steady-state experimental data.

Methods

The authors propose an end-to-end hybrid architecture called LSTM-Attention.

  • LSTM (Long Short-Term Memory): Since CEST data (Z-spectra) is sequential—consisting of signal intensities across various frequency offsets—LSTMs are used to capture the dependencies within the sequence.
  • Multi-Head Attention: The authors added an attention mechanism to help the model "focus" on specific frequency offsets that contribute most to the output, such as those corresponding to specific chemical groups (e.g., amides or hydroxyls).
  • Positional Encoding: Because standard attention mechanisms are position-insensitive, the researchers added positional encoding to ensure the model understands the order of the frequency offsets.
  • Physics-Informed Training: Instead of relying on scarce human data, they used the Bloch-McConnell equations to simulate 350,000 paired Z-spectra (short-saturation vs. long-saturation) to train the network.

Datasets

The study utilized two primary data sources:

  1. Simulated Dataset: A massive synthetic dataset of 350,000 paired Z-spectra generated using a 7-pool Bloch-McConnell model. This covered a wide range of physiological parameters (water, amides, guanidyl, hydroxyl, etc.) to ensure the model could generalize to different tissue types.
  2. In Vivo Rat Model: To evaluate the model in a real-world scenario, the researchers scanned 8-week-old rats with brain tumors using a 7T MRI. They used 1.5s saturation scans as input and 5s saturation scans as the "ground truth" reference.

Key Insights and Ideas

  • Simulation as a Bridge: The most significant "aha moment" is the successful use of physical simulations to train a model that performs well on biological tissue. This bypasses the "expensive" and often impossible task of collecting massive amounts of labeled experimental MRI data.
  • Attention to Detail: The inclusion of the attention mechanism allowed the model to outperform standard RNNs and LSTMs. It specifically improved the prediction of chemical groups in the "downfield" and "upfield" regions of the spectrum, which are often the most diagnostically relevant.
  • Pixel-by-Pixel Prediction: The framework operates on individual pixels, meaning it can reconstruct a full QUASS image by processing the Z-spectrum of every point in the MRI slice independently.

Key Results

The LSTM-Attention model was compared against five other popular architectures (MLP, RNN, LSTM, GRU, and BiLSTM) and emerged as the clear winner:

  • Linear Regression: The model achieved a coefficient of determination ($R^2$) of at least 0.9748 across all representative frequency offsets.
  • Image Quality: In terms of structural similarity (SSIM), the model scored an impressive 0.9991, indicating the predicted images were nearly identical to the long-scan references.
  • Error Reduction: The Mean Square Error (MSE) was remarkably low at $1.68 \times 10^{-4}$, outperforming the other models by an order of magnitude in several frequency ranges.

Limitations and Future Work

While the results are promising, the authors note a few areas for further development:

  • Clinical Translation: The current study was performed on a 7T small-animal scanner. Moving this to clinical 3T scanners used in hospitals will require further validation.
  • Low-Field MRI: The authors suggest that this framework would be particularly beneficial for low-field MRI, where reaching a steady state is even more challenging.
  • Specific Metabolites: Future work could investigate less visible CEST effects, such as those from guanidyl or amine groups, which are harder to detect but clinically significant.

Summary updated

Abstract

Chemical exchange saturation transfer (CEST)-magnetic resonance imaging (MRI) often takes prolonged saturation duration (Ts) and relaxation delay (Td) to reach the steady state, and yet the insufficiently long Ts and Td in actual experiments may underestimate the CEST measurement. In this study, we aimed to develop a deep learning-based model for quasi-steady-state (QUASS) prediction from non-steady-state CEST acquired in experiments, therefore overcoming the limitation of the CEST effect which needs prolonged saturation time to reach a steady state. To support network training, a multi-pool Bloch-McConnell equation was designed to derive wide-ranging simulated Z-spectra, so as to solve the problem of time and labor consumption in manual annotation work. Following this, we formulated a hybrid architecture of long short-term memory (LSTM)-Attention to improve the predictive ability. The multilayer perceptron, recurrent neural network, LSTM, gated recurrent unit, BiLSTM, and LSTM-Attention were included in comparative experiments of QUASS CEST prediction, and the best performance was obtained by the proposed LSTM-Attention model. In terms of the linear regression analysis, structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean-square error (MSE), the results of LSTM-Attention demonstrate that the coefficient of determination in the linear regression analysis was at least <i>R</i><sup>2</sup> = 0.9748 for six different representative frequency offsets, the mean values of prediction accuracies in terms of SSIM, PSNR and MSE were 0.9991, 49.6714, and 1.68 × 10<sup>-4</sup> for all frequency offsets. It was concluded that the LSTM-Attention model enabled high-quality QUASS CEST prediction.
BibTeX
@inproceedings{yang2024end,
  author = {Wei Yang and Jisheng Zou and Xuan Zhang and Yaowen Chen and Hanjing Tang and Gang Xiao and Xiaolei Zhang and Xiaolei Zhang and Xiaolei Zhang},
  title = {{An end-to-end LSTM-Attention based framework for quasi-steady-state CEST prediction}},
  booktitle = {Frontiers in Neuroscience},
  year = {2024},
  doi = {10.3389/fnins.2023.1281809},
}
Figure