Author: Lorenzo Arsini, Andrea Ciardiello, Fabio Massimo D'Amore, Stefano Giagu, Federico Giove, Carlo Mancini-Terracciano, Cecilia Voena 👨🔬
Affiliation: Istituto Superiore di Sanità, Sapienza University of Rome, Università Sapienza Roma, Magnetic Resonance for Brain Investigation Laboratory, Museo Storico della Fisica e Centro di Studi e Ricerche Enrico Fermi 🌍
Purpose: To leverage newly developed heteronuclear magnetic resonance imaging (MRI) techniques, particularly sodium (23Na) imaging, for identifying potential biomarkers of Alzheimer's disease—such as total sodium concentration (TSC) and intracellular sodium fraction (ISF)—in human brain parenchyma. Because 23Na MRI typically suffers from elevated noise levels, a deep learning–based denoising procedure was developed to enhance image quality.
Methods: A 3D residual U-Net architecture was designed and trained for denoising 3D complex k-space MRI data. The training dataset paired MRI scans from the publicly available FastMRI dataset with measured noise from a sodium scanner at the Santa Lucia Foundation Research Center. To match the sodium noise characteristics, the original 2D multi-coil FastMRI data were converted and regridded into 3D single-coil format.
Results: When trained to remove approximately 5% noise on the k-space, the 3D residual U-Net achieved, on the test set, a structural similarity index measure (SSIM) of 0.99 on denoised images, whereas the noisy images had an SSIM of 0.95. When recovering MRI images from the k-space, the improvement in SSIM becomes more evident: from 0.49 to 0.98. These findings indicate that the model effectively suppresses noise while preserving structural details and that even a small improvement on the k-space can lead to substantial gain in the real space.
Conclusion: This deep learning–based denoising approach serves as a crucial first step toward robust Na imaging for quantifying TSC and ISF in Alzheimer’s disease research. By improving image quality and reducing noise artifacts, the proposed technique may facilitate accurate biomarker identification and support early disease diagnosis.