Background:
Chemical exchange saturation transfer (CEST) MRI provides noninvasive detection of biomarkers related to myocardial creatine concentration, intracellular protein levels, and tissue fibrosis. These molecular properties offer important insights into pathologies like heart failure with preserved ejection fraction (HFpEF) [1], a pathology that remains clinically challenging to identify and is associated with increasing mortality rates. In parallel, CEST MRI of the liver provides insight into metabolic dysfunction and functional information about tumor microenvironment [2]. Despite its diagnostic potential, the long scan times associated with traditional body CEST MRI hinder its routine clinical use.

Hypothesis/Objective:
This work aims to substantially reduce scan time by developing a physics-aware vision-transformer-based deep-learning framework for rapid image reconstruction.

Methods:
A physics-aware vision transformer framework (Fig. 1) was developed to receive as input both undersampled radial k-space trajectories and the corresponding acquisition-parameter information, inspired by the recent application of a similar architecture for molecular parameter quantification [3]. The transformer was trained to reconstruct high-SNR, high-quality images that reflect an oversampled target (1000 spokes, 3.33x beyond Nyquist). Cardiac and liver CEST MRI datasets were acquired from 114 mice (271 scans total) using a 7T Bruker system, following the ungated radial FLASH protocol described in [4]. The dataset was partitioned into training (80 mice), validation (7 mice), and test (27 mice) sets. The radial trajectories were retrospectively undersampled and trained on 50, 150, and 250 spokes, corresponding to acceleration factors of R = 6.0, 2.0, and 1.2 relative to the Nyquist limit. Traditional reconstruction was performed for comparison using the BART NUFFT toolbox. Data augmentation (rotations and flips) was performed, and the loss was defined as a combination of SSIM and L1. Model selection was performed based on the minimum average Mean Squared Error (MSE) on the validation set. The model hyperparameters, including the window size and the weights of the loss function, were optimized using a Bayesian optimization framework (Optuna).

Results:
Representative cardiac and liver images, obtained using the conventional, proposed, and fully sampled reference, are available in Fig. 2, demonstrating the model efficacy in suppressing aliasing artifacts and mitigating noise across both target organs and across various frequency offsets. The neural network reconstruction significantly outperformed the undersampled NUFFT reconstruction across all acceleration factors and metrics (Fig. 3). At the highest acceleration (R=6.0), the NN reduced the NRMSE from 15.37±4.19[%] to 4.87±1.57[%] (p < 0.001). For the 150 spokes (R=2.0), the NN achieved a superior NRMSE of 3.57±1.08[%] vs 3.98±1.25[%] (p < 0.01), demonstrating high quantitative stability. PSNR at R=6.0 was increased from 16.99±2.55 dB to 26.96±2.78 dB, and SSIM from 0.29±0.08 to 0.72±0.06, with p < 0.001 for both cases.

Conclusion:
A transformer-based framework for accelerating cardiac MRI scans was developed and validated on extensive mouse data. The successful reconstruction of 6x acceleration of both liver and cardiac anatomy and molecular information highlights the model’s generalization capabilities, suggesting its potential for diverse clinical applications.

Author
Hadas Avraham
Tel Aviv University