Anatomically Guided Deep Learning Reconstruction of Accelerated Snapshot CEST MRI
- 1University of Rochester
Abstract
Purpose: To determine whether structural MRI information can improve reconstruction of highly accelerated 3D snapshot chemical exchange saturation transfer (CEST) MRI. Methods: Fully sampled brain CEST data were retrospectively undersampled at acceleration factors (AFs) of 4, 6, 8, 10, and 12. We compared reconstruction without structural information with population-level pretraining using T1-weighted (T1w), T2-weighted (T2w), or combined T1w-T2w images, as well as direct conditioning using a co-registered subject-specific T1w image. Performance was evaluated in two held-out healthy subjects using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE) across frequency offsets. Z-spectrum fidelity and derived magnetization transfer ratio asymmetry (MTRasym) maps were additionally evaluated within three-dimensional gray- and white-matter masks. Results: Subject-specific T1w conditioning consistently achieved the highest PSNR and SSIM and the lowest MAE across AFs, with greater improvements at higher acceleration. It also improved Z-spectrum agreement, particularly in white matter and combined gray- and white-matter regions across most AFs. Improvements in MTRasym were more modest and varied across tissue types and acceleration factors. Conclusion: Incorporating subject-specific structural information can improve reconstruction of highly accelerated snapshot CEST MRI, particularly at higher acceleration factors. However, improvements in reconstructed source images did not consistently translate into better preservation of derived CEST contrast, highlighting the need for reconstruction strategies that more directly incorporate CEST spectral information.
Related
- HiC-LEGO: Biologically Guided High-Resolution 3D Genome Reconstruction Preserves Chromatin Organization at Kilobase Resolution
- stEDGE enables edge-guided multiscale reconstruction of hierarchical spatial domains and transition interfaces in spatial transcriptomics
- A Multimodal Machine Learning Framework Integrating Genomic, Epidemiological, Environmental, and Clinical Imaging Data for Accelerated Measles Detection and Transmission Inference
- TLS-Sim: An Open-Source Toolkit for Individualized Transcranial Light Stimulation Modeling with Deep-Learning-Accelerated Simulation
- Machine Learning-Based Early Prediction of Neural Microelectrode Lifetime Under Accelerated Aging Conditions
BibTeX
@misc{mensah2026anatomically,
title = {{Anatomically Guided Deep Learning Reconstruction of Accelerated Snapshot CEST MRI}},
author = {E. A. Mensah and A. Faiyaz and G. Schifitto and M. N. Uddin},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.28.755205},
url = {https://www.biorxiv.org/content/10.64898/2026.09.28.755205v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Anatomically Guided Deep Learning Reconstruction of Accelerated Snapshot CEST MRI
AU - E. A. Mensah
AU - A. Faiyaz
AU - G. Schifitto
AU - M. N. Uddin
PY - 2026
DA - 2026/09/28
PB - bioRxiv
DO - 10.64898/2026.09.28.755205
UR - https://www.biorxiv.org/content/10.64898/2026.09.28.755205v1
AB - Purpose: To determine whether structural MRI information can improve reconstruction of highly accelerated 3D snapshot chemical exchange saturation transfer (CEST) MRI. Methods: Fully sampled brain CEST data were retrospectively undersampled at acceleration factors (AFs) of 4, 6, 8, 10, and 12. We compared reconstruction without structural information with population-level pretraining using T1-weighted (T1w), T2-weighted (T2w), or combined T1w-T2w images, as well as direct conditioning using a co-registered subject-specific T1w image. Performance was evaluated in two held-out healthy subjects using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE) across frequency offsets. Z-spectrum fidelity and derived magnetization transfer ratio asymmetry (MTRasym) maps were additionally evaluated within three-dimensional gray- and white-matter masks. Results: Subject-specific T1w conditioning consistently achieved the highest PSNR and SSIM and the lowest MAE across AFs, with greater improvements at higher acceleration. It also improved Z-spectrum agreement, particularly in white matter and combined gray- and white-matter regions across most AFs. Improvements in MTRasym were more modest and varied across tissue types and acceleration factors. Conclusion: Incorporating subject-specific structural information can improve reconstruction of highly accelerated snapshot CEST MRI, particularly at higher acceleration factors. However, improvements in reconstructed source images did not consistently translate into better preservation of derived CEST contrast, highlighting the need for reconstruction strategies that more directly incorporate CEST spectral information.
ER -