Introduction: Saturation transfer (ST) MRI is an increasingly investigated contrast mechanism, which enables the imaging of molecular tissue properties in living humans noninvasively. It has shown promise for a variety of clinical applications [1], including tumor detection and grading [2], early stroke characterization [3], neurodegenerative disorder imaging, and kidney disease monitoring [4]. However, the integration of ST-MRI into clinical practice has been slow and limited, due to the lengthy acquisition times required for obtaining fully quantitative tissue parameter maps. Recently, imaging techniques which combine biophysical models with artificial intelligence (AI) were suggested for accelerating quantitative ST-MRI acquisition and reconstruction [5,6]. However, the complexity of the multi-proton-pool in-vivo environment and the challenge in accurately modeling the large number of free tissue parameters significantly limit the accuracy of this approach.

Objective: The goal of this work was to develop a biophysical-model-free technique for drastically accelerating both the acquisition and reconstruction time of quantitative molecular MRI, while simultaneously extracting water relaxation and magnetic field maps.

Methods: A two-step AI pipeline was designed to (i) learn the robust human tissue response to radio-frequency excitation (Fig. 1a), and (ii) exploit it for multi-parameter tissue property mapping (Fig. 1b). First, a U-Net Transformer [7] was implemented to receive pairs of pseudo-randomly acquired ST images and their associated acquisition parameters. The transformer was trained to perform sequence-to-sequence prediction, and accurately generate a new set of contrast-weighted molecular images, in response to a user defined unseen set of acquisition parameters. Next, an expansion of the network was realized, for the simultaneous quantification of the semisolid macromolecule volume fraction and exchange rate, water T₁ and T₂, and B₀ and B₁. Validation was performed using 8 human subjects (healthy volunteers and a brain tumor patient) scanned on a 3T MRI (Prisma, Siemens Healthcare) at three different imaging sites.

Results: The acquisition time was accelerated by 80%-94% using the proposed AI pipeline compared to state-of-the-art (30 s instead of 8.5 min). The reconstruction of six fully quantitative maps for the whole brain was achieved in less than 10 s. An excellent agreement was obtained between the AI-predicted quantitative brain maps and the ground truth reference (normalized root mean squared error (NRMSE) = 5-8%, peak signal to noise ratio (PSNR) = 24-30, structural similarity index metric (SSIM) = 0.81-0.92). The AI-generated contrast-weighted images (Fig. 2) and the quantitative molecular and water brain maps were visually and perceptually similar to the reference standard output (Fig. 3), while the B₀ and B₁ maps served as a useful means for explaining the AI results and noise removal.

Conclusions: The proposed biophysical-model-free AI approach enables a drastic reduction in scan time while retaining the integrity and accuracy of quantitative MRI parameter maps. The suggested AI system enables the rapid generation of multi-contrast chemical exchange saturation transfer (CEST) and semisolid magnetization transfer (MT) images, and may easily be expanded for a wide variety of additional contrast mechanisms.

Acknowledgments: This project received funding from the European Research Council under the Horizon Europe program (grant agreement no. 101115639, BabyMagnet). This paper reflects only the author’s view, and the European Research Executive Agency is not responsible for any use that may be made of the information it contains.

Author