DEISM: Deep Reconstruction Framework With Self-Calibration Mechanisms for Accelerated Chemical Exchange Saturation Transfer Imaging

• Published in: IEEE transactions on biomedical engineering Vol. 72; no. 8; pp. 2413 - 2424
• Main Authors: Xu, Jianping, Zu, Tao, Hsu, Yi-Cheng, Wang, Xiaoli, Chan, Kannie W. Y., Zhang, Yi
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.08.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Brain - diagnostic imaging; Brain Neoplasms - diagnostic imaging; Brain tumors; Calibration; Chemical exchange saturation transfer (CEST); Coils; Correlation; Data acquisition; Deep Learning; Deism; Encoders-Decoders; Estimation; fast imaging; Humans; Image processing; Image Processing, Computer-Assisted - methods; Image quality; Image reconstruction; Imaging; Imaging techniques; learned artifact priors; Magnetic resonance imaging; Magnetic Resonance Imaging - methods; Medical imaging; Neuroimaging; Phantoms, Imaging; Redundancy; Self calibration; Sensitivity; Training
• ISSN: 0018-9294; 1558-2531; 1558-2531
• DOI: 10.1109/TBME.2025.3543403

Objective: The prolonged scan time of chemical exchange saturation transfer (CEST) imaging, caused by multiple data acquisitions over the varying saturation offset frequencies, necessitates accelerated imaging techniques. In this work, the artifact information is exploited as an important prior for CEST image reconstruction by exploring the spatial-frequential redundancy in the artifact field. Specifically, we proposed a novel deep reconstruction framework with self-calibration mechanisms (DEISM) for highly accelerated CEST imaging. DEISM features two successively concatenated structures: i) a model-based network responsible for initial image reconstruction from undersampled multi-coil k-space data, and ii) a data-driven artifact suppression (AS) network that estimates and corrects the residual artifacts in a self-calibrated manner. In addition, a novel encoder-decoder architecture with a multi-scale feature fusion mechanism is developed and utilized for robust artifact estimation and artifact correction. We trained the DEISM framework end-to-end using simulated data, and evaluated its performance on both healthy volunteers and brain tumor patients, using retrospectively or prospectively undersampled data at various acceleration factors. Experimental results demonstrated the feasibility of the data-driven AS concept and the effectiveness of exploiting the spatial-frequential correlation in the artifact field. By integrating the image artifact priors into the learning-based CEST image reconstruction process, DEISM can provide high-quality source images, molecular maps, and CEST spectra, outperforming the other conventional and state-of-the-art reconstruction techniques.