Deep neural network based CEST and AREX processing: Application in imaging a model of Alzheimer’s disease at 3 T

Abstract

Purpose: To optimize and apply deep neural network based CEST (deepCEST) and apparent exchange dependent-relaxation (deepAREX) for imaging the mouse brain with Alzheimer's disease (AD) at 3T MRI. Methods: CEST and T 1 data of central and anterior brain slices of 10 AD mice and 10 age-matched wild type (WT) mice were acquired at a 3T animal MRI scanner. The networks of deepCEST/deepAREX were optimized and trained on the WT data. The CEST/AREX contrasts of AD and WT mice predicted by the networks were analyzed and further validated by immunohistochemistry. Results: After optimization and training on CEST data of WT mice, deepCEST/ deepAREX could rapidly (~1 s) generate precise CEST and AREX results for unseen CEST data of AD mice, indicating the accuracy and generalization of the networks. Significant lower amide weighted (3.5 ppm) signal related to amyloid β-peptide (Aβ) plaque depositions, which was validated by immunohistochemistry results, was detected in both central and anterior brain slices of AD mice compared to WT mice. Decreased magnetization transfer (MT) signal was also found in AD mice especially in the anterior slice. Conclusion: DeepCEST/deepAREX could rapidly generate accurate CEST/ AREX contrasts in animal study. The well-optimized deepCEST/deepAREX have potential for AD differentiation at 3T MRI.

Relayed nuclear Overhauser effect weighted (rNOEw) imaging identifies multiple sclerosis

“…At low field strengths, the saturation power has to be low enough to avoid large DS effect. Owing to the extremely slow exchange rate (≤20 Hz), rNOE can be fully saturated even with low saturation powers (<1 μT) ( Huang et al, 2021a , Huang et al, 2021b ). This makes rNOE imaging suitable for clinical applications at 3T.…”

Lorentzian-Corrected Apparent Exchange-Dependent Relaxation (LAREX) Ω-Plot Analysis—An Adaptation for qCEST in a Multi-Pool System: Comprehensive In Silico, In Situ, and In Vivo Studies

“…Among others, Zaiss et al showed the potential for deep learning in qCEST imaging [39]. Recently, Huang et al published two fast and accurate ways to generate CEST/AREX contrast maps using DeepCEST and DeepAREX in animal experiments [40]. Based on those studies and our results, neural networks could be developed to eliminate the influence of other pools.…”

Molecular Imaging of Brain Tumors and Drug Delivery Using CEST MRI: Promises and Challenges

“…Recently, deep learning-based methods [15,99,100,104,188,190,191] have also been applied to obtain CEST contrasts to speed up and simplify the post-processing of CEST. Deep learning utilizes neural networks that are composed of multiple processing layers to extract information from data [192].…”

“…It is worth noting that the initial values and boundary values may affect the accuracy of multi-pool Lorentzian fitting and thus need to be defined properly [181]. At low field strength, some CEST peaks overlap with adjacent peaks, thus the total pool numbers need to be adjusted accordingly [99,188].…”

“…Glang et al further improved the performance of deepCEST at 3 T using a probabilistic deep learning approach which provides additional information about the reliability of the prediction [99]. In addition to deepCEST, Huang et al demonstrated that the AREX contrasts can be predicted by a deep neural network (deepAREX)which requires both Z-spectra and T1 as inputs[188]. Chen et al utilized an artificial neural network to extract CEST properties (ANNCEST), such as metabolite concentration, exchange rate, and B 0 /B 1 .…”